<?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 Max Huber on Medium]]></title>
        <description><![CDATA[Stories by Max Huber on Medium]]></description>
        <link>https://medium.com/@maxhuber?source=rss-98ca34860d7------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*_M3ZtpVbM-ZwyiDUm8Yyzg.jpeg</url>
            <title>Stories by Max Huber on Medium</title>
            <link>https://medium.com/@maxhuber?source=rss-98ca34860d7------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 05:35:37 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@maxhuber/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[Transformers for tabular data]]></title>
            <link>https://medium.com/@maxhuber/transformers-for-tabular-data-3d32cce3a48c?source=rss-98ca34860d7------2</link>
            <guid isPermaLink="false">https://medium.com/p/3d32cce3a48c</guid>
            <category><![CDATA[attention]]></category>
            <category><![CDATA[transformers]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[fraud-detection]]></category>
            <dc:creator><![CDATA[Max Huber]]></dc:creator>
            <pubDate>Tue, 17 Oct 2023 19:05:53 GMT</pubDate>
            <atom:updated>2023-10-17T19:05:53.060Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QQxZno_Mg8vHdkss" /><figcaption>Photo by <a href="https://unsplash.com/@fin21?utm_source=medium&amp;utm_medium=referral">FIN</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p><strong>Transformers have been proven to be a good architecture of Neural Networks over the last couple of years, showing state-of-the-art performance in Natural Language Processing and Computer Vision tasks. The most popular Large Language Models such as GPT or Llama are all based on Transformers and their architectures only differ in details. From generating images to having conversations, the range of possibilities with those Transformer networks is huge since they can process huge amounts of text or even videos effectively.</strong></p><p>But what about the classic tabular data we all started learning Machine Learning with?</p><p>I believe Transformers can be very powerful in use-cases with tabular data. I give you the following example:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O28IUv_MjmNuefETVfhfiQ.png" /><figcaption>In our example, a company has many vendors which each sell different products and have a different way of invoicing.</figcaption></figure><p><strong><em>A company buys products and services from its vendors. The vendors can differ in various ways. Some can sell for instance only fruits, while others might sell vegetables or cars. But each company differs also in its way of invoicing the company. So there is no general pattern in finding invoices that might be wrong. Wrong invoices might be issued twice after a while, and have a wrong amount or processing. Since each vendor has different products and internal invoicing, a suspicious invoice from Vendor 1 might be not suspicious in the context of Vendor 2. Hence, context is required.</em></strong></p><p>Context? That’s the keyword!</p><p>Since Transformers have proven their utility with contextual data, we can modify their original architecture in order to make them useful even for this use case.</p><p>Transformers use the <strong>Attention</strong> mechanism which makes them so effective on sequential data, like sentences. I recommend watching the video of Andrew Kartpathy on building GPT from scratch, he explains Attention very well (<a href="https://www.youtube.com/watch?v=kCc8FmEb1nY">https://www.youtube.com/watch?v=kCc8FmEb1nY</a>).</p><p>Attention has the advantage that every token representation (<em>key, value, query</em>) is checked against the representation of all other tokens in order to find relationships between them on both short and long distances. That’s how you usually get those <em>n_token x n_token</em> matrices which show the relationships between tokens in a sentence.</p><p>However, since the original Transformer architecture was created to solve language tasks, we have to adjust some things to make them work on tabular data like in our example:</p><h3>Positional Encodings</h3><p>The original Transformer architecture is made for sequences in which the order is important. For instance, in a sentence, you want to know if two tokens are very close to each other, or if they are the first and last tokens of a book. Both scenarios are valid and are considered in the architecture by Positional Embeddings. They are basically weights that encode the index of a token and add it to the actual token representation.</p><p>In our use case, the order of the invoices is not important. We only need to consider the dates of them, which we can encode during the data preprocessing into sine and cosine representations. Therefore, we remove the Positional Encodings.</p><h3>Decoder attention masking</h3><p>Another consequence of the importance of the token order is the attention masking in the decoder of a Transformer. For instance, when generating text, each token should only consider the preceding tokens, but not the tokens coming later. This ensures that the model cannot look into the future and use knowledge that might not be available yet.</p><p>In our use case, we do not have this requirement which is why we remove the masking in the Attention, making invoices see also future invoices.</p><h3>Categorical features</h3><p>The original Transformer uses Token embeddings in order to encode each token of a sentence. This is not applicable in our use case in which we use tabular data with non-discrete values. Discretizing our features will result in a huge loss of information, hence we will use our pre-processed data for all non-discrete values.</p><p>We also have some discrete values, such as the vendor ID or years. We can use Embeddings for those values. Hence, an embedding is being trained for every vendor, making them distinguishable.</p><p>They are then concatenated with the other features that all have to be transformed into a normalized form.</p><h3>Code</h3><p>You can check out the code for that Transformer architecture in the <em>github</em> repository: <a href="https://github.com/mhuber97/Category_Transformer/tree/main">https://github.com/mhuber97/Category_Transformer/tree/main</a></p><p>It adapted Andrew Karpathy’s <em>nanoGPT</em> implementation which you can find here: <a href="https://github.com/karpathy/nanoGPT/tree/master">https://github.com/karpathy/nanoGPT/tree/master</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3d32cce3a48c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Whats the point of Matrix decomposition?]]></title>
            <link>https://medium.com/@maxhuber/whats-the-point-of-matrix-decompositions-15435781658f?source=rss-98ca34860d7------2</link>
            <guid isPermaLink="false">https://medium.com/p/15435781658f</guid>
            <category><![CDATA[eigen-decomposition]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ml-so-good]]></category>
            <category><![CDATA[svd]]></category>
            <dc:creator><![CDATA[Max Huber]]></dc:creator>
            <pubDate>Thu, 30 Dec 2021 12:50:05 GMT</pubDate>
            <atom:updated>2022-01-04T21:03:57.190Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HfrUozVjgUvxN5-RGnIg3w.jpeg" /></figure><p>This topic was a big black hole in my head, so I tried to bring the information together that helped me understand <strong>Eigendecomposition</strong> and <strong>Singular Value Decomposition (SVD)</strong>.</p><p>For this I’m going to quickly recapitulate <strong>Determinants</strong> and <strong>Basis</strong>. If you’re already familiar with it, you can directly jump to the paragraph <strong>Eigendecomposition</strong>.</p><h3>Determinant</h3><p>A determinant answers the question of how an area scales through a certain transformation <strong><em>E</em></strong>?</p><p>For example:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sJBj2ya3WUScJ18HpqlhwQ.png" /></figure><p>So in this given example the determinant of <strong><em>E</em></strong>, or the scaling factor of a transformation, is <strong><em>-3</em></strong>. What does the negative determinant mean now? This means that the orientation of space has been inverted or the axis have been flipped. So the determinant<em> </em><strong><em>-3</em></strong> would mean that the axis are flipped over and the area is scaled by <strong><em>3</em></strong>.</p><p>An interesting property is when the determinant of a transformation is equal to <strong><em>0</em></strong>. That means that the area of the transformation would become zero because the base vectors align with each other. This scenario reduces the dimension of a vector.</p><h3>Basis</h3><p>The basis vectors of a coordinate system describe the direction of the axis in a coordinate system. For example, a two-dimensional coordinate system could be described through the basis vectors <strong><em>x</em></strong> and <strong><em>y</em></strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/966/1*72o6c5IRHGbVQH5jwerzHw.png" /></figure><p>So whenever you would describe anything in this basis, the values would be scaled with the basis vectors. It doesn’t even matter what values those basis vectors contain, so they could instead also be:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/962/1*9I-am2wXp_QiBLjQKZ01zg.png" /></figure><p>So to translate a vector<strong> <em>x_b</em></strong> from coordinate system (base) <strong><em>b</em></strong> into our coordinate system <strong><em>a</em></strong>, we can use the translation of the coordinate systems. We know how the base vectors of <strong><em>b</em></strong> are represented in base <strong><em>a</em></strong>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pfVL1_YMRb987nUlixATWA.png" /></figure><p>So to translate <strong><em>x_b</em> </strong>we multiply each basis vector <strong><em>b</em></strong> with the certain coordinate of the vector <strong><em>x</em></strong>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N-lptRb1YiqN8STc9Tb35g.png" /></figure><p>So this we have a transformation matrix that translates from basis <strong><em>b → a</em></strong>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1014/1*STSnbSUGlv67eCOaogjiqw.png" /></figure><p>In order to go the other way around, from basis <strong><em>a →b</em></strong> we need the inverse of <strong><em>T</em></strong> what is <strong><em>T⁻ ¹</em></strong> (I don’t cover how to calculate the inverse).</p><p>In order to translate a transformation <strong><em>oₐ</em></strong> from one basis <strong><em>a</em></strong> to basis <strong><em>b</em></strong> the following needs to be done:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tMWjjy9PW0uXFNW3il2GQA.png" /></figure><p>Firstly, we transform from base <strong><em>b → a</em></strong> with <strong><em>T</em></strong>. Then we apply the transformation <strong><em>oₐ</em></strong> in base a before we transform it back again into base<strong> <em>b</em></strong> with the inverse transformation <strong><em>T⁻ ¹</em></strong>. So translating the vector <strong><em>v</em></strong> would look like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/918/1*E9hLX6d9Ia77m0K8xkcP4g.png" /></figure><h3>Eigendecomposition</h3><p>The Eigenvector <strong><em>v</em></strong> is a vector for a certain Matrix transformation <strong><em>A</em></strong> that doesn’t change direction through this transformation. However, it can be stretched or squeezed by a factor that is called the Eigenvalue <strong><em>λ</em></strong>. Accordingly, the transformation <strong><em>A</em></strong> on its Eigenvalue <strong><em>v</em></strong> will give the same result as just stretching or squeezing the Eigenvector <strong><em>v</em></strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/904/1*8aJKa08vRe8XEXLL3Yyb9w.png" /></figure><p>If the Eigenvalue is equal to <strong><em>1</em></strong>, then the Eigenvector describes the rotational axis of the matrix transformation.</p><p>To calculate the Eigenvectors, we can bring the equation into the following form:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F-drxliFRr4P1-nzUwNCow.png" /></figure><p>So the product of<em> </em><strong><em>A - λI</em></strong> and <strong><em>v</em></strong> needs to be zero. A product of a transformation (<strong><em>A - λI</em></strong>) and a vector (<strong><em>v</em></strong>) can only become zero if the associated transformation <strong><em>A - λI</em></strong> squishes the vector <strong><em>v</em></strong> into a lower dimensional space — that means that the determinant (scaling factor of the area through the transformation) has to be zero! It is in lower dimensional space then, because the base vectors align into the same direction and therefore they don’t open an area (which the determinant describes).</p><p>In lower dimension vector <strong><em>v</em></strong> becomes a point instead of a vector.</p><h4>What are the Eigenvectors now good for?</h4><p>Eigenvectors can be very helpful to compute operations which are complex for matrices but easy for diagonal matrices, because they can transform matrices (transformations) into the diagonal form. In diagonal form computations, like raising a matrix to a power is much easier.</p><p>For that, the matrix has to be brought into the base vector representation of the Eigenvector, called <strong>Eigenbasis</strong>. If the matrix has enough Eigenvectors (minimum the number of axis), the translation will guarantee a <strong>diagonal matrix </strong>representation where complex calculations can be applied. Then the result can easily be brought back into the origin coordinate system. The diagonal matrix contains the Eigenvalues on the diagonal, the scaling factor of the particular axis in Eigenbasis.</p><h4>Summary</h4><ul><li>Bringing a transformation into the Eigenbasis:</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QEplnq0xEo6y41_nKzPB4w.png" /></figure><ul><li>Bringing a diagonal matrix in the Eigenbasis to the original basis:</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o_YhDNDUkSmTFcyB1K2Giw.png" /></figure><h4>Example</h4><p>To make use of this, here an example:</p><p>We have a transformation matrix <strong><em>A</em></strong>, on which we want to apply an operation, for example raising it to the power <strong><em>b</em></strong>. Since this operation is complex for non-diagonal matrices we want to bring the diagonal form. For this we need the Eigenvectors <strong><em>p</em></strong> which build the columns of matrix <strong><em>P</em></strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eDoqF55LHtFwc5ZXj1pNNA.png" /></figure><p>We have brought the <strong><em>A</em></strong> into the Eigenbasis, where this transformation is illustrated through a diagonal matrix <strong><em>D</em></strong>.</p><p>Now, we can apply our intended operation to <strong><em>D</em></strong> and bring it back into our origin-base:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DC4oPfn9t3M1phGiQjYYyw.png" /></figure><p>That’s all!</p><h3>Singular Value Decomposition (SVD)</h3><p>Why is there another form of decomposition? Isn’t the Eigendecomposition enough?</p><p>Unfortunately not! The Eigendecomposition is only applicable for squared matrices. Singular Value Decomposition addresses all types of matrices, and always exists!</p><p>But the concept stays the same:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*29XVdeSXIgPNcKC4B3Yj1g.png" /></figure><p>We have a transformation <strong><em>A</em></strong> that can be split into three operations that bring the transformation into another basis where a squishing or stretching takes place. In contrast to the Eigendecomposition we don’t have one other basis but two. The reason for this is the given dimensionality of the origin matrix <strong><em>A</em></strong>. Instead of being just squared it can be in the form of <strong><em>m×n</em></strong>.</p><p>Firstly, <strong><em>V^T</em></strong> of size <strong><em>(n×n)</em></strong> brings a vector from basis <strong><em>A</em></strong> into basis <strong><em>B</em></strong>.</p><p>Then <strong><em>Σ</em></strong> of size <strong><em>(m×n)</em></strong> like <strong><em>A</em></strong> performs a scaling on the basis vectors, since it is a diagonal matrix. However, it does not only applies a transformation but also transforms the calculation from basis <strong><em>B</em></strong> to basis <strong><em>C</em></strong>.</p><p>Finally, <strong><em>U</em></strong> of size<strong> <em>(m×m)</em></strong> brings the calculation back from basis <strong><em>C</em></strong> to basis <strong><em>A</em></strong>.</p><p>Eigendecomposition and Singular Value Decomposition are deeply connected with each other. In the following we can see why.</p><h4>Reminder</h4><p>We know that for squared matrices <strong><em>A ∈</em> ℝ<em>^{n×n}</em></strong> there always exist a matrix <strong><em>P</em></strong> which can transform the matrix into the Eigenbasis.</p><p>A matrix <strong><em>A</em></strong> can be transformed into a squared matrix by being multiplied with the inverse <strong><em>A^{-T}</em></strong>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HrOgxOcHa9jAb98kJ9225g.png" /></figure><h4>SVD</h4><p>What we need to find out for the SVD are the matrix <strong><em>V</em></strong> of shape <strong><em>n×n</em></strong> and <strong><em>U</em></strong> of shape <strong><em>m×m</em></strong>. So we start with <strong><em>V, </em></strong>the right-singular vectors, since <strong><em>V</em></strong> is the right part of the equation <strong><em>A = UΣ V^T</em></strong> and illustrates the <strong><em>n×n</em></strong> basis. Accordingly, we perform the left-singular vectors by computing <strong><em>U</em></strong> of shape <strong><em>m×m</em></strong>.</p><h4>Right-singular vectors <em>V</em></h4><p>This knowledge helps us to bring a rectangular matrix <strong><em>A </em></strong>of shape <strong><em>n×n</em></strong> into a squared form which then allows us to apply Eigendecomposition:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2hpL4IrEuvJEpwnbZb0iqQ.png" /></figure><p>Decomposing it with the Singular Value Decomposition equation with can also receive the following:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m1u6-ajv5RgA7lSfevBKiA.png" /></figure><p>Those two formulas look very similar and lead to the assumption:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SFAmKB91eBm7zOCTT1_a_A.png" /></figure><p>Accordingly, we can apply the Eigendecomposition from <strong><em>A^{T}A</em></strong> to compute the Eigenvectormatrix <strong><em>P</em></strong> which equals to <strong><em>V</em></strong>.</p><h4>Left-singular vectors U</h4><p>Similarly to the right-singular vectors, we want to obtain a matrix of shape <strong><em>n×n</em></strong> what we can get through <strong><em>AA^T</em></strong>.</p><p>Accordingly, we can compute the Eigenvectors which form the matrix <strong><em>P</em></strong> which equals <strong><em>U</em></strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mU0jkIU-_698hPyqMLDoSA.png" /></figure><p>For this, we can find the Eigenvectors in U.</p><h4>Singular values <strong>Σ</strong></h4><p>In order to obtain the singular values <strong><em>Σ</em></strong>, we have to take the already computed diagonal matrix <strong><em>D (Σ^{T}Σ)</em></strong> from the right-singular or left-singular vectors and take the square-root. This is easy, because <strong><em>D</em></strong> is a diagonal matrix that contains the Eigenvalues on the diagonal.</p><p><strong><em>Σ</em></strong> is of the shape as <strong><em>A</em></strong>, and if <strong><em>m≠n</em></strong> the bigger axis is padded with zeros.</p><p>For example:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CJq3bvBaqAXH271op_8A8Q.png" /></figure><h4>What is it good for?</h4><p>Through Singular Value Decomposition important information can be derived. The singular values in <strong><em>Σ</em></strong> tell how discriminant each value of the matrix is. So if a value is high, that means that the certain dimension is important for the data representation. Consequently, if it is low the dimension does not carry a lot of information.</p><p>How does this help us?</p><p>This helps us to identify dimensions which are not important for the whole picture. We can for example remove the dimensions which don’t contain important information to compress the data.</p><h3>Conclusion</h3><p>The key take-away of Eigendecomposition and Singular Value Decomposition are that you can transform a matrix into another basis to either make complex operations easy to compute (Eigendecomposition) or compress data by only regarding the most meaningful datapoints (SVD).</p><p><a href="https://medium.com/mlearning-ai/mlearning-ai-submission-suggestions-b51e2b130bfb">Mlearning.ai Submission Suggestions</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=15435781658f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The 7 lessons I learned on my first hut hike]]></title>
            <link>https://medium.com/@maxhuber/the-7-lessons-i-learned-on-my-first-hut-hike-f29a28358879?source=rss-98ca34860d7------2</link>
            <guid isPermaLink="false">https://medium.com/p/f29a28358879</guid>
            <category><![CDATA[mountains]]></category>
            <category><![CDATA[hiking]]></category>
            <category><![CDATA[alps]]></category>
            <category><![CDATA[climbing]]></category>
            <category><![CDATA[hut]]></category>
            <dc:creator><![CDATA[Max Huber]]></dc:creator>
            <pubDate>Wed, 11 Aug 2021 15:18:18 GMT</pubDate>
            <atom:updated>2021-08-11T15:18:18.364Z</atom:updated>
            <content:encoded><![CDATA[<p>This summer I went on my first hut hike in the Stubai Alps in Austria and here are 7 lessons that I learned from it.</p><figure><img alt="mountain view" src="https://cdn-images-1.medium.com/max/1024/1*MnoCiNvBObg3A7xrfSPf4Q.jpeg" /><figcaption>mountain view</figcaption></figure><h3>1. The food is great!</h3><p>Austrian food is amazing! There are so many specialties no matter if you’re a more salty or more sweet food lover. Arriving at the hut after a cold day in the mountains you can start with a hot soup which will warm you up from inside-out. Usually, the huts offer half board meals which include a three/four course menu in the evening and a breakfast. The menu in the evening always starts with a soup and a salad, followed by a meal you can choose after your arrival. There is always a vegetarian option which is not self-evident considering all the meaty traditional meals in Austria. If you’re lucky you visit a hut in which they are offering seconds (you definitely get enough food there). As a dessert they often serve cake or berry-yogurt.</p><p>To get enough energy for your next hike you can get as much food as you want at breakfast. The hut always offers cereals with yogurt or fresh milk and bread with cheese and sausage. Sometimes they also prepare egg-meals and fruits or vegetables.</p><h3>2. People are open for a conversation or card games</h3><p>I would consider myself as a city-person so I am used to anonymous behaviour in bars or restaurants. The hut-life was completely different. After arriving from your hike you usually don’t crawl into your room but stay in the public room of the hut.</p><p>It’s quite easy to get into conversation with anyone there and chat about your tour or anything. There are many interesting people there and you have the opportunity to meet new ones every day.</p><p>I also recommend bringing some card games with you. It’s usually not a big additional weight for your backpack and it is often a good opener to get into conversation with people.</p><h3>3. You don‘t pay for the accommodation but for the location and its circumstances</h3><p>We were a little surprised when we received the receipt at check-out. We paid about 70–80€ p.p. per night. That’s what we roughly paid the night in the valley before getting up to the huts. But comparing the bedrooms, the bathrooms etc. the hut obviously could not compete.</p><p>But what you have to keep in mind is the location. The huts (often) don’t have any infrastructure around like roads or water pipes etc. Anything you get there is brought there by people in a very elaborate way. You will probably forget the money when you have the beautiful morning view of the mountains in the morning.</p><h3>4. Don‘t underestimate the altitude and start slowly.</h3><p>My friends and I are living in low altitude cities. So the first tour at about 2.000m — 2.700m was unusual for our bodies. We didn’t want to accept it at first, but our bodies had to get used to this height. 3 out of 4 of us felt a little weird after the first tour and even got some headache. It settled after the night but I would not recommend you to exaggerate at your first tour and go to even higher places if you’re not used to the air up there.</p><h3>5. Functional shirts will let you carry much less</h3><p>I bought two functional shirts in advance of our hut hike and was sceptical at first. I was used to the basic sport shirts and wondered how much better these functional shirts could be.</p><p>But already after our first tour I was totally convinced of them. I was sweating a lot but these shirts dried so fast that I had no chance of getting cold at all. At the same time it’s also great for the basic layer inside an onion-look. These 15€ were totally worth it!</p><h3>6. If the temperature isn’t about to change a lot, a thunderstorm is not likely to come.</h3><p>We were quite sceptical about the weather conditions in the alps. It’s said that the weather can change quickly when you’re up on a mountain. The weather forecast was pretty bad and we were afraid to get into a thunderstorm. When we revealed our concerns to some locals, they reassured us that a thunderstorm would only appear after a large temperature change. Otherwise, we would only get some rain.</p><p>Ultimately, the weather forecast is always worse than it actually gets. But don’t risk too much! ;)</p><h3>7. Bags are saving time and keeping order</h3><p>You actually don’t need to take much with you at a hut hike. But there are some clothes you need, like underwear and socks, some shirts, sweaters, food, etc. There are some bags which have multiple openings in it, so you can get you things from above and from below and some that don’t offer that.</p><p>Nevertheless, if you’re on your hike and need to get something from the inside of your backpack it can get quite messy. You cannot keep order in the inside, so the solution for this is: small bags in which you can collect the individual type of stuff. So for example: one bag for underwear, socks and shirts and another one for sweaters and jackets and a third for food. The main advantage of this is that you always keep your things inside the certain bag in order. So no more mess inside your backpack!</p><figure><img alt="morning view from the hut" src="https://cdn-images-1.medium.com/max/1024/1*OSuktCgZJYSgueKDD4M3UA.jpeg" /><figcaption>view from the hut room in the morning</figcaption></figure><h3>Conclusion</h3><p>These are the 7 lessons I learned from my first hut hike. Do you have anything to add? I would appreciate your tips and lessons learned in the comments!</p><p>Greets, Max</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f29a28358879" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>