<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Raja Sekhar on Medium]]></title>
        <description><![CDATA[Stories by Raja Sekhar on Medium]]></description>
        <link>https://medium.com/@eunoiaraj?source=rss-7619cb28fc29------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*p1augVb7cwtmmJvX.</url>
            <title>Stories by Raja Sekhar on Medium</title>
            <link>https://medium.com/@eunoiaraj?source=rss-7619cb28fc29------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 13:07:53 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@eunoiaraj/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[From Basic Automation to True AI Agents: A Four-Layer Framework]]></title>
            <link>https://medium.com/@eunoiaraj/from-basic-automation-to-true-ai-agents-a-four-layer-framework-ed3edb7b5d4d?source=rss-7619cb28fc29------2</link>
            <guid isPermaLink="false">https://medium.com/p/ed3edb7b5d4d</guid>
            <category><![CDATA[definition]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[framework]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Raja Sekhar]]></dc:creator>
            <pubDate>Tue, 10 Jun 2025 13:22:54 GMT</pubDate>
            <atom:updated>2025-06-10T13:22:54.652Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Separating marketing hype from real autonomy — and a call for community feedback</em></p><blockquote><em>“Agentic AI” is 2025’s favourite buzz-phrase, yet many so-called agents are little more than rule scripts or LLM wrappers. A clear yard-stick helps architects, investors, and regulators decide whether a system is truly capable. This four-layer framework is my current attempt; I invite you to poke holes in it so we can iterate toward a sharper definition.</em></blockquote><h3>1. Core (non-negotiable) properties</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/814/1*kbfJNvMf6QiClWpGIGKI0g.png" /></figure><blockquote>Miss one of these and you likely have a tool, not an agent.</blockquote><h3>2. Capabilities that turn a basic agent into an AI agent</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/855/1*tuImHl_ZGIiT4Phu0Osy8Q.png" /></figure><h3>3. Qualities expected in modern production-grade agents</h3><ul><li><strong>Explainability &amp; Traceability</strong> — Surface the reasoning path; replay decisions.</li><li><strong>Safety / Alignment Guards</strong> — Policy checks, constrained decoding, red-team simulators.</li><li><strong>Multi-modal Competence</strong> — Handle text, images, audio, and structured data seamlessly.</li><li><strong>Lifecycle Hooks</strong> — Versioning, continuous evaluation, dev-test-prod promotion.</li></ul><blockquote><em>These aren’t strictly required for the </em>definition<em> of an AI agent, but real-world deployments collapse without them.</em></blockquote><h3>4. Quick Checklist — Spot-the-Agent in 30 Seconds</h3><p>Tick ✔️ if the answer is “yes”.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/823/1*_bgJDmC4ILwhWHQ7yigVqw.png" /></figure><ul><li><strong>0–3</strong> ✔️ → Tool / Script / Classic Automation</li><li><strong>4–5</strong> ✔️ → Intelligent Agent Candidate</li><li><strong>6–7</strong> ✔️ → Production-grade AI Agent</li></ul><h3>How to Use This Framework</h3><ol><li><strong>Architects</strong>: Map your system components to the four layers to reveal missing pieces (often memory or safety guards).</li><li><strong>Investors</strong>: Ask founders to walk through the checklist; hype evaporates quickly under concrete questions.</li><li><strong>Researchers</strong>: Propose metrics for each layer (e.g., autonomy latency, learning rate, alignment loss) to make comparisons fair.</li><li><strong>Regulators</strong>: Anchor policy discussions on which layers trigger which obligations.</li></ol><h3>Let’s Evolve the Definition Together</h3><p><em>Is explainability core or optional? Should multi-agent coordination be its own layer?</em> Share your critiques and real-world counter-examples in the comments.</p><p>I’ll compile the best insights into a v2.0 post with full attribution. Thanks for helping sharpen the collective vocabulary around AI agents!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ed3edb7b5d4d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From LLMs to AGI: Understanding AI Agents, MCP, and the Emerging World Model]]></title>
            <link>https://medium.com/@eunoiaraj/from-llms-to-agi-understanding-ai-agents-mcp-and-the-emerging-world-model-45ea53afb38e?source=rss-7619cb28fc29------2</link>
            <guid isPermaLink="false">https://medium.com/p/45ea53afb38e</guid>
            <category><![CDATA[world-models]]></category>
            <category><![CDATA[agi]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[mcps]]></category>
            <category><![CDATA[genai]]></category>
            <dc:creator><![CDATA[Raja Sekhar]]></dc:creator>
            <pubDate>Sun, 06 Apr 2025 14:45:51 GMT</pubDate>
            <atom:updated>2025-04-09T03:01:33.837Z</atom:updated>
            <content:encoded><![CDATA[<p>As large language models (LLMs) become increasingly central to modern AI systems, the demand for more general-purpose and extensible architectures continues to grow. In this post, we’ll walk through the evolution from LLMs and fine-tuning, to agentic frameworks, and the role of the new MCP (Model-Client Protocol) in standardising model-tool integration.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/533/1*fiFhpZhvikxeGrgcPD374w.png" /></figure><h3>Knowledge</h3><h4>LLMs as Knowledge Dumps</h4><p>LLMs (Large Language Models) are AI models trained on massive language datasets. They serve as statistical knowledge bases capable of generating relevant outputs from natural language inputs.</p><p>A base model without fine-tuning can perform:</p><ol><li>Sentence completions</li><li>Text generation (such as essays, code, or stories)</li><li>Translations (if trained on multilingual data)</li></ol><p>However, these models lack task awareness. They do not inherently understand or follow instructions; instead, they predict text based on patterns in their training data.</p><h3>Wisdom</h3><h4>Fine-Tuning for Purpose and Instruction</h4><p>To achieve task-specific behavior, base LLMs are fine-tuned using instruction-response pairs. This process teaches them how to respond to structured prompts.</p><p>For example, if a user asks, “Summarize this paragraph,” a fine-tuned model understands the instruction and delivers a summary, rather than continuing the paragraph.</p><p>Fine-tuned models can be:</p><ol><li>Instruction-following (e.g., summarization, translation)</li><li>Conversational (e.g., ChatGPT)</li><li>Domain-specific (e.g., legal or medical applications)</li></ol><p>Some systems use multiple fine-tuned models to handle different tasks or domains more effectively.</p><h3>Generalization</h3><h4>Two Approaches Toward Generalization</h4><p>There are two primary strategies for advancing toward general-purpose AI systems:</p><ol><li><strong>Train and fine-tune multiple models</strong> to cover a broad range of tasks and domains. This explains the current proliferation of specialized LLMs in the ecosystem.</li><li><strong>Extend existing models</strong> with external capabilities, allowing them to perform beyond what their training and fine-tuning alone enable.</li></ol><p>First approach, Fine-tuning is resource-intensive and not scalable for every potential use case. Each new requirement would demand its own model or training pipeline, which limits flexibility and slows down innovation.</p><p>The second approach is proving to be more scalable and flexible, especially when paired with the ability to access tools and data in real time.</p><h4>Extending Model Capabilities with Tools and Data</h4><p>To enable dynamic behaviors, modern systems integrate LLMs with:</p><p><strong>External data sources</strong>, providing knowledge beyond the training set.</p><p><strong>Tools</strong>, allowing models to perform actions such as API calls, database lookups, code execution, and more.</p><p>These tools act as external functions that models can invoke when needed. This design allows the model to go beyond response generation and participate in reasoning, planning, and execution.</p><h4>Agent Frameworks and the Challenge of Integration</h4><p>To support these extended capabilities, a number of orchestration frameworks have emerged, including:</p><p>LangChain, CrewAI, OpenAI Operator, AutoGen and others</p><p>These frameworks allow LLMs to:</p><ol><li>Chain multi-step reasoning processes</li><li>Invoke tools or plugins</li><li>Integrate with APIs, vector databases, and external systems</li></ol><p>However, each framework defines its own way of managing integrations. There is no standardized protocol to govern how models communicate with tools, leading to fragmentation and limited interoperability.</p><h4>Introducing MCP: The Model-Client Protocol</h4><p>MCP (Model-Client Protocol), proposed by Anthropic, addresses this gap by providing a standardized client-server protocol for model-tool interaction.</p><p>MCP offers:</p><ol><li>A <strong>model-independent</strong> interface for invoking external tools and systems.</li><li>A <strong>standardized format</strong> for structured requests and responses.</li><li>A <strong>vendor-neutral</strong> approach that allows frameworks and tools to interoperate seamlessly.</li></ol><p>This protocol allows models to be extended dynamically — without being tightly coupled to specific orchestration frameworks or integration methods.</p><h4>How MCP Works</h4><p>The following is a simplified overview of the MCP interaction pattern:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/976/0*n3zZ8yzC3SGajQw3.png" /></figure><h4>Interaction Steps:</h4><ol><li>The model is fine-tuned to recognize when it needs assistance (e.g., fetching weather, running a calculation).</li><li>It outputs a structured and generalized request that does not depend on any particular system.</li><li>The MCP Client captures this output, identifies the task, and routes it to the appropriate tool via the MCP Server.</li><li>The MCP Server executes the request through its integrations.</li><li>The response is returned to the client, which passes it back to the model.</li><li>The model uses this additional context to generate a final output for the user.</li></ol><p>This interaction can be <strong>iterative</strong>, allowing the model to engage in multi-step reasoning with intermediate tool calls.</p><h4>How MCP Differs from Agent Frameworks</h4><p>MCP is not intended to replace frameworks like LangChain or CrewAI. Instead, it complements them by acting as a <strong>common protocol layer</strong>.</p><p>While LangChain orchestrates the flow of logic and tool usage, MCP standardizes how those tools are accessed and interacted with — making it easier to build portable, modular AI systems that can work across platforms.</p><p>It’s still early in MCP’s evolution, and it remains to be seen whether it will become the dominant standard. But it clearly addresses a key limitation in today’s agentic systems.</p><h3>World Model</h3><h4>Looking Ahead: Toward Runtime Intelligence and AGI</h4><p>As we move further toward Artificial General Intelligence (AGI), the current architecture — where models rely on pre-defined tools, integrations, and data sources via MCP Servers — will continue to evolve.</p><p>In the near future, <strong>the generation of tools, system connectors, and data access modules will no longer be pre-configured</strong>, but instead dynamically created <strong>at runtime</strong> by models themselves. These <strong>MCP Server capabilities will become composable and adaptive</strong>, constructed based on the unique purpose or task identified by the parent model.</p><p>This means:</p><ul><li>Tools will not be statically defined, but instead <strong>generated on demand</strong>.</li><li>Integration logic will be inferred and constructed <strong>contextually</strong>, based on the user’s goal.</li><li><strong>External data fetching mechanisms will be tailored dynamically</strong>, using auxiliary models that assist in discovering, building, or even coding these interactions in real time.</li></ul><p>Initially, this dynamic approach may be <strong>slow or computationally expensive</strong>, but over time, we can expect:</p><ul><li><strong>Optimized processes</strong> for tool generation and reuse</li><li><strong>Community or model-built repositories</strong> of reusable tool templates, integrations, and APIs</li><li>More <strong>performant infrastructure</strong> to support this level of flexibility and autonomy</li></ul><p>If this trajectory holds — and these capabilities become <strong>fast, generalizable, and self-improving</strong> — we will be witnessing the foundation of a <strong>world model architecture</strong>, capable of adapting to any real-world problem. At that point, AGI is no longer a theoretical destination, but a functional reality.</p><h3>Conclusion</h3><p>We are witnessing a progression in LLM capabilities:</p><p>Level 1 — From models that generate text,</p><p>Level 2 — To models that follow instructions,</p><p>Level 3 — To models that can reason, act, and integrate with tools and systems.</p><p>Level 4— To models that can self improve, adapt, and evolve to be more efficent.</p><p>As we move towards more capable and flexible AI systems, <strong>generalization, extensibility, and standardization</strong> will be key. MCP offers a promising step forward in enabling model-agnostic, tool-driven intelligence at scale.</p><p>Additional Read: personal thoughts on World Model,</p><blockquote><strong><em>The World Model: Toward Adaptive, Evolving Intelligence</em></strong></blockquote><p>If models evolve to dynamically generate tools, construct system connectors, and fetch or synthesize knowledge at runtime — without being limited to pre-configured capabilities — we begin to approach what many refer to as the <strong>World Model</strong>.</p><p>A world model is more than just a large language model with external tools. It is a system that can:</p><ul><li>Understand complex goals.</li><li>Analyze and decompose tasks into smaller subtasks.</li><li>Assemble the required tools, data access mechanisms, or reasoning pathways on demand.</li><li>Interact with external systems, APIs, and even other models, in an adaptive, autonomous way.</li></ul><p>This vision moves us away from building narrow-purpose models or rigid pipelines. Instead, we imagine models that:</p><ul><li><strong>Generate runtime infrastructure</strong> by invoking specialized helper models or code generators.</li><li><strong>Build and maintain evolving knowledge bases</strong>, learning continuously from interactions with the world.</li><li><strong>Reason about reasoning itself</strong> — deciding how to approach a problem, when to seek external help, and when to act independently.</li><li><strong>Coordinate with other agents</strong>, forming dynamic networks of cognition that can collaborate across tasks or domains.</li></ul><p>If such systems become performant, fast, and efficient — capable of real-time decision-making and execution — we are no longer just interfacing with intelligent tools. We are witnessing the emergence of <strong>adaptive, world-aware intelligence</strong>.</p><p>At that point, AGI may not arrive in the form of a singular monolithic model, but rather as an ecosystem of self-organizing, self-extending models that perceive, reason, build, and evolve.</p><p>This is the future we are steadily moving toward — and MCP-like protocols, runtime tool generation, and agentic orchestration are laying the foundation for it.</p><h3>Resources to Dive Deeper</h3><ol><li><strong>Anthropic’s Model-Client Protocol (MCP)</strong></li></ol><p><em>Official Proposal &amp; Technical Explanation</em></p><p><a href="https://docs.anthropic.com/mcp">https://docs.anthropic.com/mcp</a></p><p><strong>2. The Bitter Lesson — Rich Sutton (2019)</strong></p><p><em>A foundational piece on why generality and learning matter more than hand-coded systems.</em></p><p><a href="http://www.incompleteideas.net/IncIdeas/BitterLesson.html">http://www.incompleteideas.net/IncIdeas/BitterLesson.html</a></p><p><strong>3. Self-Instruct: Aligning Language Models with Self-Generated Instructions (2022)</strong></p><p><em>Introduced instruction tuning — a key milestone in the evolution from base models to agentic models.</em></p><p><a href="https://arxiv.org/abs/2212.10560">https://arxiv.org/abs/2212.10560</a></p><p><strong>4. Toolformer: Language Models Can Teach Themselves to Use Tools (Meta AI, 2023)</strong></p><p><em>Demonstrates how LLMs can autonomously learn to call tools — aligns directly with your AGI vision.</em></p><p><a href="https://arxiv.org/abs/2302.04761">https://arxiv.org/abs/2302.04761</a></p><p><strong>5. ReAct: Synergizing Reasoning and Acting in Language Models (2022)</strong></p><p><em>Describes the method of combining reasoning with tool use — widely used in LangChain and CrewAI.</em></p><p><a href="https://arxiv.org/abs/2210.03629">https://arxiv.org/abs/2210.03629</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=45ea53afb38e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gradient Decent Optimiser evolution]]></title>
            <link>https://medium.com/@eunoiaraj/gradient-decent-optimiser-evolution-60f5a5ec4507?source=rss-7619cb28fc29------2</link>
            <guid isPermaLink="false">https://medium.com/p/60f5a5ec4507</guid>
            <category><![CDATA[aiml]]></category>
            <category><![CDATA[gradient-descent]]></category>
            <dc:creator><![CDATA[Raja Sekhar]]></dc:creator>
            <pubDate>Wed, 24 Jul 2024 13:10:03 GMT</pubDate>
            <atom:updated>2024-07-27T19:24:48.295Z</atom:updated>
            <content:encoded><![CDATA[<p>I love Gradient decent , without which our AI Models either become a guess work or it will be brute force to calculate all possibilities of weights in a range and we are still not sure if they are correct.</p><p>I am assuming you already know what are weights, Loss, Cost, model equation etc, if not, please learn/revise them and come back to this blog.</p><p>Types of Gradient Decent Optimisers covered in this blog</p><ul><li><strong>Batch Gradient Descent (BGD) (1847)</strong></li><li><strong>Stochastic Gradient Descent (SGD) (1951)</strong></li><li><strong>Mini-Batch Gradient Descent (1985-1990)</strong></li><li><strong>Momentum</strong> (1983)</li><li><strong>Nesterov Accelerated Gradient (NAG)</strong> (1983)</li><li><strong>Adagrad (Adaptive Gradient Algorithm)</strong> (2011)</li><li><strong>RMSprop (Root Mean Square Propagation)</strong> (2012)</li><li><strong>Adadelta</strong> (2012)</li><li><strong>Adam (Adaptive Moment Estimation)</strong> (2014)</li><li><strong>Nadam (Nesterov-accelerated Adaptive Moment Estimation)</strong> (2016)</li><li><strong>AdaMax</strong> (2014, as part of Adam paper)</li><li><strong>AMSGrad</strong> (2018)</li></ul><p>Ok , Let us start looking at the journey on how they evolved.</p><ol><li>Batch Gradient Descent (BGD) : Gradient descent application , where each decent step is calculated basing on all provided dataset (xi,yi). As the training datasets are huge the computation will take more time. But it guarantees convergence to global optimum (in scope).</li></ol><p>J(θ)=(1/2m) ∑i=1tom ​(hθ​(x(i))−y(i))^2</p><ul><li>m is the number of training examples.</li><li>hθ​(x(i)) is the hypothesis, defined as hθ(x)==θTx.</li><li>x(i) is the feature vector of the i-th training example.</li><li>y(i) is the actual output for the i-th training example.</li><li>θ is the vector of model parameters.</li></ul><p>If weights are considered as Theta,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/525/1*dAmpHzJ1xbotzvlzlumYIQ.png" /></figure><p>∇θ​J(θ)=1/m ​∑i=1tom​ ((hθ​(x(i))−y(i))x(i))</p><blockquote>It is actually as partial derivate by theta (theta1, theta2, ..)</blockquote><blockquote>Partial derivative of (1/2m) ∑i=1tom ​(<strong>hθ​(x(i))</strong>−y(i))² with respect to theta θ</blockquote><blockquote>theta term in bold , but as it is f(<strong>h(</strong>θ<strong>)</strong>)^n , derivative will be n*f(<strong>h(</strong>θ<strong>))</strong>)^n-1 * (derivate of <strong>h(</strong>θ<strong>) by </strong>θ)</blockquote><blockquote>= (1/2m) *2 * ∑i=1tom ​(<strong>hθ​(x(i))</strong>−y(i)) ^ (2–1) * (derivate of <strong>hθ​(x(i))</strong>−y(i) <strong>by </strong>θ)</blockquote><blockquote>considering <strong>hθ​(x(i))= </strong>∑i=1ton θ(i)*x(i) [where n is number of features]</blockquote><blockquote>= (1/2m) *2 * ∑i=1tom ​((<strong>hθ​(x(i))</strong>−y(i)) * x(i))</blockquote><p>Replacing in above equation,</p><p>θ:=θ−α(1/m ​∑i=1tom ​((hθ​(x(i))−y(i))x(i)))</p><p>So we start with initial model <strong>hθ​(x(i)) </strong>and update the θ in a loop until we reach a possible global optimal cost value.</p><p>2. <strong>Stochastic Gradient Descent (SGD) </strong>— as BGD<strong> </strong>is computationally complex , imagine you have 1 million record in the training dataset , you need to run through these 1 million input values to calculate each step until you reach optimal value needs huge compute. So in SGD they calculate new parameters for each input , instead of all inputs.</p><p>Above computation is all the same removing ∑i=1tom</p><p>θ:=θ−α * ( ​((hθ​(x(i))−y(i)) * x(i)))</p><p>this is faster for sure, but the path and probability of reaching gloabl optimal is not great.</p><p>3. <strong>Mini-Batch Gradient Descent — </strong>BGD is too complex to compute and SGD is to vague with respect to each step, so mini batch is to iterate through certain size of inputs as s, so the computations becomes as below,</p><p>θ:=θ−α(1/s * ​∑i=1tos ​((hθ​(x(i))−y(i))x(i)))</p><p>so the input m is broken into m/s mini batches and gradient descent operated on each batch at a time instead of handling the whole batch.</p><p>if we divide m datasets into B1, B2, B3…Bx batches where B1=B2=..=Bx=1 for SGD and B1=B2=..=Bx==s for mini batch of size s.</p><p>consider for 5 iterations the change in theta will be as below,</p><p>θ0​=0</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/525/1*zvRfutj-IfH9M-gbquqvdw.png" /></figure><p>4. <strong>Momentum Gradient descent</strong> — Batch Gradient descent is the best at reaching the solution compared to SGD and mini batch , but the computation is hard, so how can we achieve batch solution reducing the compute.</p><p>Batch has m loops to achieve new theta, SGD has 1 loop to reach theta (m iterations to cover the complete input m dataset) and mini batch has s loops to reach new theta (m/s iterations to cover the complete input m dataset).</p><p>so if you observe batch gradient is combination of these x iterations of theta change in SGD and Mini-batch. if we can combine impact of each batch to the next one like below,</p><p>Every step we will take the gradient from current step and some part of gradients from previous steps as below the combination of these 2 will be read as v , additional variable to remember the previous value of change,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/279/1*mN5eRraYrNYWkVYMubnLaw.png" /></figure><p>here β defines the percentage of the previous value we want to use for the next step it range is usually between 0.5 to 0.9</p><p>putting above equations together it will be like this ,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/666/1*5phb2Rx0j3dMV5NHRrYjbg.png" /></figure><p>As β is less than 1 and we are multiplying derivative by β to the power of iterations, so as we go to higher iterations the lower gradients will become negligible like β=0.6 and β²⁰ = 0.0000365</p><p>4. <strong>Nesterov Accelerated Gradient (NAG) — </strong>In momentum , we calculate each step taking one step per iterations , where as in NAG , we take 2 steps and for next calculation to see the future state if we go in same direction as the previous batch</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/666/1*qzRb51gWidweF-QPysQtsw.png" /></figure><p>For Gradient calculation, instead of using ,</p><p>θt−1 = θt−2 - vt−1​</p><blockquote>NAG uses,</blockquote><blockquote>θlookahed = θt−1 -βvt−1​</blockquote><blockquote>θlookahed = (θt−2 -vt−1) -βvt−1</blockquote><blockquote>θlookahed = θt−2 -(1+β)*vt−1</blockquote><p>so i =n summary instead of using β as multiplication factor it will use (1+β) that is the difference , not that 1+β is greater than 1 , the gradient will not be lost because of multiplication with a very small number as in previous example now it will be if β = 0.6 1+β²⁰ = 12089.258, this in other way brings the problem of exploding the gradients and making them too big a number compared to the latest gradient.</p><p>5. <strong>Adagrad (Adaptive Gradient Algorithm) — </strong>Adagrad is where we divide the learning rate by square root of sum of squares of all gradients till that step , so the learning rate will decrease for every iteration, let us look at the equations,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/391/1*XrVXVmbyMhJFeKFefUKp2w.png" /></figure><p>so the iterations work as below, a small constant epsilon is added to sum of squares to avoid zero division</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/687/1*s06IXZDd1UX0ibs6RvRQ-g.png" /></figure><p>but as you see , the G value increases as the steps progress and will reduce learning rate and in turn get it close to 0 and learning may stop early and learning will also go very slow after certain iterations as learning rate reduces.</p><p>6. <strong>RMSprop (Root Mean Square Propagation) — </strong>In Adagrad as the value of sum of squares of gradients is increasing , RMSprop started multiplying them with a hyper parameter to consider only part of the gradient squares as shown in the below iterations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/423/1*Xl9YGxzyULLUdJOsltMf1w.png" /></figure><p>Iteration on how the Gradient accumulation will vary is as below,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/747/1*qw21pQ1DCDiuvGpJKUFetQ.png" /></figure><p>as we can see the gradient values initial values are diminishing by iteration as β is usually 0.5 to 0.9 , for example, β=0.6 and β²⁰(1-β) = 0.00000365 multiplying gradient square by this will make it negligible. Now that the values are going below 0 depending on gradient values, the learning rate may diminish or increase, we don’t have control over it.</p><p>7. <strong>Adadelta — </strong>instead of using one denominator root square of gradient , it also considers the theta values change from previous steps in numerator as shown in the below equation,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/543/1*WkQ93ThT8fa6JVxcrhEAnw.png" /></figure><p>So it is change change it is sum squares of theta change terms / sum of squares of gradients with some decay on each.</p><p>Iterations work like below and the gradient term is present in both numerator and denominator with different multiples.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/714/1*wmL4olZSZTFqRqwkzuxe1A.png" /></figure><p>8. <strong>Adam (Adaptive Moment Estimation) — </strong>Adam instead of using a squared terms for both numerator and denominator for the multiplier of learning , it will use numerator without square and denominator with square root of sum of squares, as in below equation, so in this scenario the learning rate decay is much better and adaptive as per the gradient movement.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/531/1*kTgWWpqwdtdqvkxOC7ISew.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/531/1*40QxL5qL_KS7ALq0Lbxueg.png" /></figure><p>we also have a control over numerator and denominator terms to vary as the constant multipliers are different in Adam for both of them as seen above. also dividing the v and m for every iteration will help in decay of these values as the iteration increase</p><p>the way the β1 and β2 will vary basing on iterations as below, this is an example for one term, considering β1​=0.9 and β2=0.999</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/531/1*0OjIuDtTQ4yIPFFnMSPOyg.png" /></figure><p>9. <strong>Nadam (Nesterov-accelerated Adaptive Moment Estimation) — </strong>NADAM is combination of Adam and NAG , where we use Theta lookahead for cost from previous step instead of using theta(t-1) in the equation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/560/1*fyJnV9oFirAq5urgs0EcOg.png" /></figure><p>iterations for few steps will be like below,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/560/1*y6X8ToEK_okRVZ7Vh1Qusw.png" /></figure><p>mt = (1-β1)*∇θ​J(θt-1) + ​β1(1-β1)*∇θ​J(θt-2)+β1²(1-β1)*∇θ​J(θt-3)……+β1^t-2(1-β1)*∇θ​J(θ1)+β1^t-1(1-β1)*∇θ​J(0)</p><p>similar has same equation using squares of gradients, so the objective is still the same that the part of the gradient from previous calculations will be used for deriving the next theta.</p><p>The derivation of the look ahead as</p><p>θ(lookahead) = θ(t-1)+β1*m(t-1) this equation is not clear, need to look into this derivation.</p><p>10. <strong>AdaMax — </strong>taking adam equation mt and vt , vt which is basing on square of gradient is replaced with ut which takes the maximum gradient value of the past or the present like below</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/501/1*Q9arDW5XTsFLnuLxJ_BNAQ.png" /></figure><p>ut= max(β2​*ut−1​,∣g(t-1)∣)</p><p>= max(β2 * ​(max(β2​ * (β2*​ut−2​,∣g(t-3)∣),∣g(t-2)∣),∣g(t-1)∣)</p><p>as you see ut is max of the gradient approx as the β2 multiplication is applied we can’t say direct max.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/501/1*AbLhJ-GIGMKd6E0cRcjgMA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/501/1*CP8XCBTEN6JTZQ_FOMfcBQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/561/1*TIs6fouFGM9rAR-z7AfW9A.png" /></figure><p>11. <strong>AMSGrad — </strong>similar change as Adamax on Adam optimiser, but the second momentum is considered as is of Adam and the max values of the second momentum (combination of previous gradient squares + present gradient square)is taken instead of taking the max of gradient directly</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/561/1*ysr3pd4BnGaQhooylSUjpQ.png" /></figure><p>​</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/509/1*ZQVARYqNhrk00sfXLuv_ZQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/543/1*74qBq2YRGIEJ9nz-7VT2Gw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/543/1*O_KXF1lPL5TmUNDKda9KVg.png" /></figure><p>So this is how the variants of Gradient descents will work till now the performance of these gradients on specific model implementation is shown in the graph below, but the performance depends on the data how sparse is it and other factors</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/552/0*isKeNEozntOTm4Zc.png" /></figure><p>Ref: Paper — <a href="https://production-media.paperswithcode.com/methods/Screen_Shot_2020-05-28_at_6.25.01_PM_Nke4eXS.png">https://production-media.paperswithcode.com/methods/Screen_Shot_2020-05-28_at_6.25.01_PM_Nke4eXS.png</a></p><p>Tabular format of how Gradient descent evolved and what improvement they have done on the previous one,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/822/1*i87fCXVbybPCJaHJnGahcQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/822/1*DVMKyRl-zDRzZqjy7DIZag.png" /></figure><p>Thanks for reading and understanding. Have a great learning :) :)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=60f5a5ec4507" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Decision Tree — Anti Intelligence Algorithm]]></title>
            <link>https://medium.com/@eunoiaraj/decision-tree-anti-intelligence-algorithm-e710233e1416?source=rss-7619cb28fc29------2</link>
            <guid isPermaLink="false">https://medium.com/p/e710233e1416</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[algorithms]]></category>
            <category><![CDATA[decision-tree]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[learning]]></category>
            <dc:creator><![CDATA[Raja Sekhar]]></dc:creator>
            <pubDate>Wed, 06 Mar 2024 14:06:34 GMT</pubDate>
            <atom:updated>2024-03-06T14:06:34.802Z</atom:updated>
            <content:encoded><![CDATA[<h3>Decision Tree — Anti Intelligence Algorithm</h3><p>Enough of drawing these lines to predict the magic , I give up !!!</p><p>So, you’re stuck with this massive dataset, and your old tricks aren’t cutting it anymore. Linear Regression? Logistic Regression? Nah, it’s time to level up.</p><p>Enter the Decision Tree — the secret weapon for taking down big problems one step at a time.</p><p>Imagine your dataset as an army, with each feature as a different type of soldier. Your mission? Divide and conquer. Split the army smartly to increase your chances of success.</p><p>But how do you decide where to split? That’s where entropy comes into play.</p><p>Entropy is like a measure of chaos in your dataset. The higher the entropy, the more confused you are. Your goal is to reduce this confusion and make clearer decisions.</p><p>Here’s how it works: You start by picking a feature to split your army on. Let’s say you choose “Soldiers on Horses” If this split makes things clearer — like separating soldiers on flat ground versus riding horses — then you’ve reduced the chaos, or entropy. You’re closer to victory.</p><p>But the game isn’t over after one split. You keep checking the entropy at each stage. If it’s still high, you find another feature to split on, making things even clearer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/831/1*xYUVUbJulL-uLR31c6Et3A.png" /></figure><p>you remember logistic regression , we calculate loss by entropy , the same entropy can be used here as well for each Node</p><p><em>H</em>(<em>S</em>) = −∑(<em>i</em>=1to<em>c)</em>​ P<em>i</em>​ * log2​(P<em>i</em>​)</p><p>and for each split we calculate either Information gain or Gini Index, this shows how much we are reducing the entropy the lossing probability, we make a choice basing on the war and features we are fighting,</p><p><em>IG</em>(<em>S</em>)=<em>H</em>(<em>S</em>)−∑ size(<em>Sc)/size(S)</em>​×<em>H</em>(<em>Sc</em>​)</p><p>Where S is Parent , Sc are its Childs, H(S) is Entropy for each node,</p><p>With each split, you’re evaluating the probability of winning. You’re making smarter decisions, reducing chaos, and increasing your chances of success at every turn.</p><p>So, the next time you’re faced with a daunting dataset, remember the power of Divide and Conquer with the Decision Tree. It’s all about making strategic moves, reducing chaos, and cracking the code one step at a time.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e710233e1416" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Dr NN (Neural Networks) is trying to cure Covid (example for Backpropagation)]]></title>
            <link>https://medium.com/@eunoiaraj/dr-nn-neural-networks-is-trying-to-cure-covid-example-for-backpropagation-e05e74ad5f87?source=rss-7619cb28fc29------2</link>
            <guid isPermaLink="false">https://medium.com/p/e05e74ad5f87</guid>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[backpropagation]]></category>
            <category><![CDATA[learning]]></category>
            <dc:creator><![CDATA[Raja Sekhar]]></dc:creator>
            <pubDate>Mon, 24 Apr 2023 12:02:07 GMT</pubDate>
            <atom:updated>2023-04-24T12:15:11.938Z</atom:updated>
            <content:encoded><![CDATA[<p>Before I get started on this, Thanks for your support and appreciation, If you are new here, please go through my previous blogs, so that you can understand this better.</p><ol><li><a href="https://medium.com/@eunoiaraj/yes-you-can-understand-machine-learning-c60fd9fbc182?source=your_stories_page-------------------------------------">Yes, You can understand Machine Learning</a></li><li><a href="https://medium.com/@eunoiaraj/is-my-machine-learning-well-dfe7b7a0feb8?source=your_stories_page-------------------------------------">Is my Machine learning well?</a></li><li><a href="https://medium.com/@eunoiaraj/neural-networks-are-a-war-formation-of-linear-regs-with-special-powers-62b038d1f5df?source=your_stories_page-------------------------------------">Neural Networks are a war formation of Linear Regs with special powers !!</a></li></ol><p>I am using Covid vaccine as an example in this blog to explain how Neural networks learn and provide predictions. Please join the fictional journey,</p><p>There is Covid everywhere, people are suffering and researchers started looking at the pathogen on how it attacks, what is virus genetic structure, how it impacts human body parts, and other details, considering these as the inputs (X features) for our Dr Neural network, let us try to see how he would create a Covid vaccine as a preventive cure.</p><p>As per initial data analysis , he understood, that this needs expertise from different researchers/scientists in different studies and it will be a compound of different mixtures which will be solved by different scientists. Even though these experts can solve the problem in parts and provide respective substances to cure specific symptoms / impacts of the virus, Dr NN needs to understand how he can put them together to produce the Covid cure which works.</p><p>DR NN distributed the work across 3 labs with Lab1 10, Lab2 15 and Lab3 5 experts at each lab respectively, they would create a solution of their own and pass it to the next lab, the next lab takes all the individual substances, and mixes them in certain portion, as their solution and pass it to the next lab. After passing through 3 labs it comes to Dr NN to come up with the final compound for the vaccine and provide it for trials.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xdxj67oS-R9hLcTYZIaRyQ.png" /></figure><p>It would be a miracle if that worked as there is a lot of guesstimate with respect to how each substance needs to be added to make the final vaccine. But on the other hand, people are suffering a lot and DR NN was curiously waiting for the results. Finally, the results came back, but it didn’t work, they also gave details as what is the impact, virus behavior, change in symptoms, and other complications.</p><p>Dr NN looks at this trial response and his expected cure response, he puts together the differences and analysis on how he should change the weightage(W4) of the substance(A3) passed from Lab3, so that it could be better.</p><p>Post that he would send this trail response information along with his analysis and planned changes to the previous lab Lab3, the scientists at Lab3 will process this info and come up with how much they need to change their mixture weightage(W3) of each compounds they receive from the previous stage as per expectations of Dr NN‘s compound and the final vaccine impact differences. Later they pass these details to Lab2 , where they would repeat the same process adjust their weightage(W2) and later pass it to Lab1 and finally Lab1 would also adjust their weightage(W1) accordingly.</p><blockquote>Dr NN →. Lab3 →. Lab2 →. Lab1</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m7K_JCqHv6nSYCzh06Dm9Q.png" /></figure><p>after everyone adjusted their strategy on how they wanted to mix the medicine, we would run the vaccine development process from Lab1 to Dr NN again and trigger Clinical trials. This exercise will keep going until Dr NN (with help of these labs) will try to create a successful vaccine for Covid.</p><p>We have observed the clinical trials is giving us the amount of error between expected and produced (or predicted) solution, Dr NN and 3 labs tries to look at the amount of change that their substance mixture weights with respect to the amount of change on the final error, so that they can get the expected output to next immediate lab and also for final prediction(proper Covid vaccine).</p><pre>Lab1 Weight = W1 (w11, w12,..,w10)<br>Lab2 Weight = W2 (w21, w22,…,w215)<br>Lab3 Weight = W3 (w31, w32,…,w35) <br>DRNN Weight = W4 (w41)</pre><p>If the difference between expected Vaccine output Y and Predicted output is Yhat is considered as E (Error), they weightage W based on the difference in E with respect to difference in W, dE/dW</p><pre>Wx (new) = Wx - dE/dWx, (x = 1, 2, 3, 4 of above weights)</pre><p>We should understand that dWx depends and should be according to dWx+1, as what gets into Vaccine is the compound of all these substances as per weights at different labs + DrNN.</p><blockquote>dW1 depends on dW2 depends on dW3 depends on dW4</blockquote><pre>dE/dW1   ~   dE/dW2  ~   dE/dW3   ~   dE/dW4</pre><p>This way of going backward from the Output layer Dr NN to Lab1 and adjusting the weights in a way to reduce the error between predicted and actual output values is called Backpropagation in Neural networks.</p><p>Now the next challenge in front of DR NN is Covid variants would you create a new NN of layers or would use the neural network with changes in the output layer, what would you do? Please add in the comments. <em>Finally, the Nobel award goes to Dr NN for discovering the fictional Covid vaccine.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e05e74ad5f87" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Neural Networks are a war formation of Linear Regs with special powers !!]]></title>
            <link>https://medium.com/@eunoiaraj/neural-networks-are-a-war-formation-of-linear-regs-with-special-powers-62b038d1f5df?source=rss-7619cb28fc29------2</link>
            <guid isPermaLink="false">https://medium.com/p/62b038d1f5df</guid>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Raja Sekhar]]></dc:creator>
            <pubDate>Mon, 17 Apr 2023 04:18:13 GMT</pubDate>
            <atom:updated>2023-04-17T04:18:13.226Z</atom:updated>
            <content:encoded><![CDATA[<h3>Neural Networks are a war formation of Linear Regs with special powers !!</h3><p>Neural Networks are simple, it is like drawing more lines on your data stacked at multiple levels.</p><p>Before I get into the details, please understand when I say a line as in linear regression, it is not necessarily a line as y = w*x + b, Once x has more than one feature it will go multi-dimensional, but here I am assuming it to be a line for better understanding.</p><p>In linear regression for the given dataset, we draw a line that could help us predict y, If you see in a Neural network, we have neurons or units arranged in multiple layers to traverse from the input to the output layer.</p><p>Each Neuron works with a combination of the linear equation z = w*x + b and an activation function g(z), as per our understanding of the linear regression, we are drawing a line on the input points. Activation function is applied on top of it, but let us focus on the linear equation as the lines are drawn on the given points at each layer.</p><p>Initially, you would provide the input layer with features x1 to xn, these are given to input hidden layer 1 h1 with l1 units, each of these l1 units draw a line on these points as shown below, correct?</p><figure><img alt="lines drawn on the given input points , ideally these won’t be lines basing on the number of input features." src="https://cdn-images-1.medium.com/max/1024/1*hSN9cZg7JWRCAejG_2yIig.png" /></figure><p>For understanding purposes, I have shown lines in this image, but it would be a different figure based on the number of input features which will add multiple dimensions to this graph. In this graph I had l1=8, so there are 8 lines drawn on the input data. Activation function let us say Relu (Rectified linear unit g(z)=max(0,z)) will strip this line’s negative part with values less than 0 and show us the positive piece(which is still a line).</p><p>Now let us move to the second hidden layer h2, if you see the input points A1 are nothing but the points on the h1 layer lines predicted by hidden layer 1. If the hidden layer has l2 units, these l2 units now will draw lines on top of the h1 layer in a way that these lines try to be close to the points on A1 lines, that is the above line points.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-OxaBBAT20bz4Vpgb7VHXg.png" /></figure><p>so if you imagine a multi-storied building,</p><p>0 — Ground floor is the input data X,y and</p><p>1 — First floor has l1 lines drawn by h1 hidden layer , as output A1 points (after applying Relu on top of Z1=W1*X+b1 (X = A0 = Input))</p><p>2 — Second floor has l2 lines drawn by the h2 hidden layer, as output A2 points (after applying Relu on top of Z2=W2*A1+b2 )), so second floor is trying to draw a line which has a minimum error for points on the first floor lines.</p><p>3 — Third floor has l3 lines drawn by the h3 hidden layer same as above, these lines target to minimize the distance for points on A2 previous floor lines.</p><p>after L floors as per the NN architecture for the specific problem statement, we have L-1 floors of lines drawn on top of the input data on the ground floor. the key takeaway or understanding is each layer treats the previous layer’s output as its features and tries to fit into the model to build a solution on top of it.</p><p>PH — top floor or the penthouse or the output layer considers the output of the L-1 floor points on the lines and draws a line or lines based on the problem statement as binary classification, multi-class classification, regression, etc we apply respective activation functions on the ZL.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/970/1*houl52vGRPvDJPsQ8T5CNQ.png" /><figcaption>Picture Credits to datascience-enthusiast</figcaption></figure><p>As in the above diagram, you are putting a stack of lines on top of the input data in multiple layers, where each line tries to reduce overall loss for the points of the previous layer and produce the closest output as required.</p><blockquote>Xn = 34 features A0 → l1 = 8 (draw 8 lines on A0 get A1) → l2 = 15 ( draw 15 lines on A1 get A2) → ……… → L=1 (draw single line on top of L-1 lines points and apply logistic regression for binary classification)</blockquote><p>If the activation function is different from Relu the points may not be a straight line (tanh, signmoid, etc), it could be in a shape as per the activation function , but they are a set of points on which we draw the lines (Zn) in the next layer.</p><p>So that is how we set up soldiers of linear regression on different floors to attack the data and the objective is to win the war by the time they reach the top floor PH with a solution as required. Start your war, Peace.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=62b038d1f5df" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Is my Machine learning well?]]></title>
            <link>https://medium.com/@eunoiaraj/is-my-machine-learning-well-dfe7b7a0feb8?source=rss-7619cb28fc29------2</link>
            <guid isPermaLink="false">https://medium.com/p/dfe7b7a0feb8</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[cost-function]]></category>
            <category><![CDATA[mathematics]]></category>
            <dc:creator><![CDATA[Raja Sekhar]]></dc:creator>
            <pubDate>Mon, 27 Mar 2023 08:28:54 GMT</pubDate>
            <atom:updated>2023-03-27T08:28:54.790Z</atom:updated>
            <content:encoded><![CDATA[<p>Machine is learning on how to respond to a given problem data to produce a solution, As Machine doesn’t know about these problems and solutions, we provide training to Machine for learning (like companies provide to freshers or college to students), during training the machine might not be able to give best results directly, so we need to assess how good is the machine output for the training data (X,y). So we see the difference between the expected output y and machine produced output yhat like this yhat-y which shows the loss for a given training example, As we need to verify it across training data, we do a Sum for n training dataset,</p><pre>Σ(i=1 to n) (yhati-yi)</pre><p>as the difference could be positive or negative, we will go for either absolute value or square</p><pre>Σ(i=1 to n) (yhati-yi)^2</pre><p>as this sum varies based on the no of training examples, we will take a mean (or average), so that it is not influenced by n. (divided by 2 for easy math, but not sure of the specific reason, if you know, please add in the comments)</p><pre>(1/2m) * Σ(i=1 to n) (yhati-yi)^2</pre><p>This value is called the cost function, as it shows what it cost us as a loss if we use a specific machine learnt model. It is denoted by J. Our target during training will be to reduce this cost so that overall loss is minimal for expected output.</p><p>This is good, but this raises more questions, how do I know if my cost is high, or low and how do I make changes to the model and reach an optimal value? Good, if you are thinking like this, let us cover that detail in the next blog.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dfe7b7a0feb8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Yes, You can understand Machine Learning]]></title>
            <link>https://medium.com/@eunoiaraj/yes-you-can-understand-machine-learning-c60fd9fbc182?source=rss-7619cb28fc29------2</link>
            <guid isPermaLink="false">https://medium.com/p/c60fd9fbc182</guid>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Raja Sekhar]]></dc:creator>
            <pubDate>Tue, 21 Mar 2023 04:35:07 GMT</pubDate>
            <atom:updated>2023-03-22T13:50:52.634Z</atom:updated>
            <content:encoded><![CDATA[<p>Everything in this world can be represented as data, including problems and solutions; when you look at problem data and solution data, there will be some kind of relation between these datasets, this relation helps us in solving the problem. This might sound simple, but very important.</p><p>Each problem data has a pattern or patterns, few of these patterns will help us in finding the solution data. Before we go further let us look at some examples of problem data and solution data, houses data — house prize data, patients data — cancer status data, CCTV footage data — person identification data, vehicle data — speeding data, student marks data — admission approval data, etc. <br>As you look at these examples, the relation and need for deriving a solution is understood, but how to arrive at a solution, given a problem is not straightforward, it depends on multiple features within the problem data and identifying the patterns relevant to the required solution, the patterns could vary based on the solution data requirement, like house data to house prize or house data to maintenance cost per year or house data to owner happiness. As you notice, few of the patterns and relation to a solution are obvious and directly understandable, but few are not. <br>Different features and patterns of problem data(like age, size, parking space, owners, groundwater, fire safety, etc.) will have a different impact on the solution, depending on the required solution (house price, maintenance cost, owner happiness, etc.)</p><p>Are you clear till here? Problem data, Problem features, Problem data patterns, Feature and pattern weightage in connection with Solution data. If you still don’t understand these terms, please go through it again. And there is nothing wrong with it.</p><p>Machine Learning models are an attempt to create a mathematical equation that will fit the given problem data X(input) in a way they can predict a solution y(output) for future inputs. Key point is that we believe all these X features work together in a combination with some weight and correlation to produce this y and we need to find that relation or function between X variable(s) and y variable(s). Depending on the data complexity and output requirement (as Regression, Classification(Binary, and Multiclass) we would pick different degrees of polynomial or mathematical function which would fit the data. It could go complex by using multiple equations or methods in combination (either single or multi-level or layers) to align with the required prediction.</p><p>These equations can have multiple inputs such as x1(size), x2(age), x3(groundwater), x4(parking), etc, these in ML are called features and the equation aligns with a different degree, combination, and weightage to these features and tries to predict a given y with minimum error. this equation or combination of equations is called an ML Model. <br>ML prediction is where we apply input problem data to this equation and get a solution(prediction), If the output is a continuous number it is regression output(120K, 110K, 250K, 100K) and if the output is defined set of values it is classification ([good or bad],[orange or white or green],[malignant or benign], etc.)</p><p>let us put all this into an example, the problem data could be list of house details as x1(area), x2(bedrooms), x3(bathrooms), x4(stories) and solution data (prediction) y will be (price).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/683/1*8bf7oWgHjjG_ruNZHgxy7w.png" /></figure><p>As you see X and y has a relation or pattern , if the x1area is increasing the y price is increasing , the x2 no of bed rooms are increasing the y price is increasing , but in some cases the x1 area is increasing , but the x2 bedrooms are lower , y price is not increasing and x3 bathrooms doesn’t impact y price as much as x1 area or x2 bedrooms., similarly each of x1, x2, x3 , x4 individually has some correlation with y and combination of them has different impact on y. let us say if we measure how much each feature or combination of feature impacts y price, we can name them , w1 for x1, w2 for x2, w3 for x3 , w4 for x4 , w5 for x1x2, w6 for x2x3, w7 for x1x2x3x4 etc. so y could be written as = w1x1+w2x2+w3x3+w4x4+w5x1x2 ….</p><p>so this <strong>equation </strong>to get the solution y using X (as x1,x2,x3,x4) can be called one of the ML model between given problem and required solution.</p><p>If we understand this, the next questions will be, <br>How to find the features? How to find the correct weightage? How to find the correct one? How to find the error? how to predict the values? but let us hold them for now, let us celebrate that we understood what is machine learning and we can explain it to anyone. Give it a shot in the comments.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c60fd9fbc182" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Every Child needs to read this …]]></title>
            <link>https://medium.com/@eunoiaraj/every-child-needs-to-read-this-34112d2cda1f?source=rss-7619cb28fc29------2</link>
            <guid isPermaLink="false">https://medium.com/p/34112d2cda1f</guid>
            <category><![CDATA[pain]]></category>
            <category><![CDATA[love]]></category>
            <category><![CDATA[fiction]]></category>
            <category><![CDATA[short-story]]></category>
            <dc:creator><![CDATA[Raja Sekhar]]></dc:creator>
            <pubDate>Mon, 13 Feb 2017 09:34:57 GMT</pubDate>
            <atom:updated>2017-02-14T09:40:43.531Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BKWYfCEq0-djMfRgFshMng.jpeg" /></figure><p>Her brain is occupied with a sense of joy and a feeling of scary pain. As if a feeling of a student desperately waiting to leave the school, but scared of the entrance exam.</p><p>Everyone is tensed, the phone calls have been coming to get the status. She knew that she is not sick, but she still needs to sleep on a sick bed. <br>There is one more person, anxiously walking to and fro in the hallway, Feeling that he needs to face a situation, which he has never experienced. He needs to support her, soothe her in pain, but whatever he does, he can’t experience, take or reduce her pain.</p><p>While the world takes different injections as a cure, she is the only one takes it for pain. The whole environment is scary. She recollects the numerous advice from everyone. Slowly she can feel the pain going through her nerves Creating tremors in her brain. She is afraid, but she wants to bare it because all that matters for her is the result.</p><p>Unexpectedly the pain multiplied in minutes, she suddenly felt like she was hit by a heavy vehicle. People gathered around to soothe her and keep her calm. Her eyes looked for one person, he mattered the most than anyone else. She believes he can reduce the pain, he does some magic and makes everything alright.</p><p>Walked into the room, back of his mind, he is very puzzled, he never knew, how it feels, what to do ?, what decision to take ?. But he knew he is her confidence. She believes in him more than anything else. he can’t let her down, so he walked in with great confidence. Took her close to him and said, it’s tough, but we need to go through it, don’t worry everything will be alright.<br>She didn’t reply, all she did was catch his hand as if it’s a magic wand and that would make her better. He can feel her pain by the way she was holding his hand. As if she is falling down from a tall building.</p><p>He kept telling her, what she has been telling about this situation and reminded how she wanted it. As the words are flowing the pain is growing. She literally couldn’t take it anymore, she is kicking the bed with her legs and shaking her hands, the pain is literally shaking her up.<br>Someone ran to call for help.</p><p>Every minute is hell, she started shouting, every time she shouts, his confidence is coming down, he has become helpless, he could do nothing, but see her suffer in really horrible pain.</p><p>Helper, people around, everybody are trying to control her, she caught hold of his shirt and pulled it really hard out of pain, he is losing control, his helplessness is breaking its limits and converting into water in his eyes. But he knows that is the last thing she expects, a single drop from his eye &amp; she would start feeling more pain. As every minute passes, the things are getting much worse.</p><p>He was standing there holding her close to him, feeling her pain. Her pain reached the limits, she felt like her bones are breaking,<br>She couldn’t take it anymore, eyes filled with water, she looked into his eyes and nodded her head as if she can’t. She didn’t speak a word, but he heard it, <em>“Plz, help”</em>.</p><p>There were no words, he just kept her back on bed, he said give me one min and i will set everything. There were many people around trying to stop him. He requested them to be silent, he knew what he is doing and that is the right thing.</p><p>He made one call and said, “ I want it to be done and right now”.<br>A person who always speaks in requested tone has suddenly demanded &amp; the other person understood&amp;reached in 2 mins. As soon as the person was there, they shifted her, she is still shouting, he was stopped from going inside, but the way she was shouting, he ran inside pushing everyone, he requested her to bare it for 5 mins and everything will be alright. <br>He knew she is trying her 100%,</p><p>but he couldn’t find better words, she was quickly rushed in and he was sent out. There were many people outside talking, asking questions etc, but he couldn’t listen to anything, the way she shouted is still resounding in his ears. He fell deep in her tears and couldn’t come out of it <br>In 5 mins, there was a cry inside, this time it is different, everyone one is excited, someone came from inside and told its a girl. A moment, 3 people were born, a baby, a mother, and a father.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fYsZ6CxWDcgz6TKSZu6yAg.jpeg" /></figure><blockquote><strong>Love u mom.</strong></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=34112d2cda1f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Prevention is easy, Yes it is !!]]></title>
            <link>https://medium.com/@eunoiaraj/prevention-is-easy-yes-it-is-8e732588b05f?source=rss-7619cb28fc29------2</link>
            <guid isPermaLink="false">https://medium.com/p/8e732588b05f</guid>
            <category><![CDATA[health]]></category>
            <category><![CDATA[happy]]></category>
            <category><![CDATA[healthsaathi]]></category>
            <category><![CDATA[lifestyle]]></category>
            <category><![CDATA[preventive]]></category>
            <dc:creator><![CDATA[Raja Sekhar]]></dc:creator>
            <pubDate>Tue, 31 Jan 2017 12:42:27 GMT</pubDate>
            <atom:updated>2017-01-31T12:48:49.188Z</atom:updated>
            <content:encoded><![CDATA[<p>Today I want to tell you a small story, This story is about Rahul, Rahul is 25 year old guy working for a software company and recently got a new bike. It’s a big day in office, Rahul is expecting a promotion. He quickly started from home, he needs to take 2nd right turn in VIP road for his office. While he is crossing the first right, a biker on the left side took a sudden right turn, Rahul didn’t even notice him, Rahul’s bike ran into the other bike, Rahul fell down at a speed and his head hits the footpath…</p><p>I would like to pause the story here, I would like to talk about 3 things,</p><p>First Rahul should have <strong>awareness, </strong>not just that he needs to wear a helmet, but an accident can happen any time and he needs to be cautious about it.</p><p>Second, he should have <strong>assessed, </strong>Is he crossing a vehicle ?, Is that biker showing a right signal ?, Is there any turn ?</p><p>Third, A <strong>preventive measure</strong>, if any of the above is true, Rahul needs to slow down , observe the vehicle and pass it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/825/1*1RN04AjItbENZ_L8mzW73g.jpeg" /></figure><p>Taking the story a little back, Rahul started from home on his new bike, wearing his new helmet, he started to office. He needs to take 2nd right, he noticed a biker giving right signal, he slowed down, let the other bike turn and reached office. He got a surprise welcome in office from his manager with a promotion letter as Team lead with good hike.</p><p>Preventive Healthcare starts with that <strong>basic awareness</strong> about you and your family health, the easy ways in which you can <strong>track &amp; access</strong> the health status and take few simple <strong>preventive measures, </strong>which gives you, healthy and happy family.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/851/1*CclXANetziAwl6P0R_DeFg.jpeg" /></figure><p>As you see in the story, Preventive healthcare is not costly checkups, treatments, operations or hospital visits, they are simple, very affordable, easy lifestyle changes. For example, do you know maintaining food timings daily will boost your metabolism and help you maintain your weight ? So next time, don’t keep weight loss as your new year agenda, you should pick one lifestyle solution as your agenda and trust me 100% you will see a difference.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8e732588b05f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>