<?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 Fernando Jesus Garcia Hipola on Medium]]></title>
        <description><![CDATA[Stories by Fernando Jesus Garcia Hipola on Medium]]></description>
        <link>https://medium.com/@fhipol?source=rss-29cd67b1dbf8------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*7TQOb-JwnIqsXS58hAKCoQ.jpeg</url>
            <title>Stories by Fernando Jesus Garcia Hipola on Medium</title>
            <link>https://medium.com/@fhipol?source=rss-29cd67b1dbf8------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 12:11:37 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@fhipol/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[Are Words Abstract Representations Encoded within LLMs?]]></title>
            <link>https://medium.com/@fhipol/are-words-abstract-representations-encoded-within-llms-9bd8038b6a7e?source=rss-29cd67b1dbf8------2</link>
            <guid isPermaLink="false">https://medium.com/p/9bd8038b6a7e</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cognitive-science]]></category>
            <category><![CDATA[language]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Fernando Jesus Garcia Hipola]]></dc:creator>
            <pubDate>Sat, 05 Jul 2025 09:41:02 GMT</pubDate>
            <atom:updated>2025-07-05T09:41:02.985Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fw9SqYOxAwXhJZz0ZdyMcg.jpeg" /><figcaption>Abstractness: Photo by <a href="https://unsplash.com/@aznbokchoy?utm_content=creditCopyText&amp;utm_medium=referral&amp;utm_source=unsplash">Lucas K</a> on <a href="https://unsplash.com/photos/blue-and-orange-smoke-wQLAGv4_OYs?utm_content=creditCopyText&amp;utm_medium=referral&amp;utm_source=unsplash">Unsplash</a></figcaption></figure><p>In this article, I am introducing my master’s thesis research, where I explored how information related to word abstractness is encoded internally in a Large Language Model (LLM).</p><p>First of all, I need to mention a paper which greatly inspired me: <em>Do Large Language Models represent internally space and time?</em> from<a href="https://arxiv.org/abs/2310.02207"> Gurnee and Tegmark (2023)</a>, where they analyzed if LLMs create spatial and time representations in their internal activations.</p><iframe src="https://cdn.embedly.com/widgets/media.html?type=text%2Fhtml&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;schema=twitter&amp;url=https%3A//x.com/wesg52/status/1709551516577902782&amp;image=" width="500" height="281" frameborder="0" scrolling="no"><a href="https://medium.com/media/79597386ad57d164a412435a5840ccd8/href">https://medium.com/media/79597386ad57d164a412435a5840ccd8/href</a></iframe><p>To tackle such a question, the paper’s authors arranged datasets about spatial entities (<em>i.e.</em> cities and their coordinates), and time entities (<em>i.e.</em> historical figures and their birth dates). The entities names were included on text templates used to prompt diverse LLMs, and then the authors managed to predict the time/spatial data solely from the internal model activation by a technique called <strong>probing</strong>.</p><p>Basically a<strong><em> probe</em></strong> is just a separate simpler model (<em>i.e.</em> a regression model) trained on the LLM hidden activations to predict other target variables. In the case of that paper, the variables were the spatial and time coordinates information on their datasets. Note that only associated entities names were inserted into the prompts, but the spatial and time coordinates were not indeed shared directly to the LLMs! As some of the linear probe models predicted successfully the spatial/time data from the internal LLM activations, suggesting this information is encoded linearly within the researched LLMs.</p><h4><strong>From Representationalism to Abstract Meaning</strong></h4><p>The mentioned worked by Gurnee and Tegmark falls within the domain of <strong>AI interpretability</strong>, a field of study to understand the internal mechanisms of deep learning models. It contributes to a central debate questioning if LLMs are merely sophisticated pattern matchers — often so-called “<em>stochastic parrots</em>” — , or if they do develop meaningful internal representations. The latter view, aligning with a form of <strong>representationalism</strong>, suggests that LLMs build internal models to effectively process and understand information about the world.</p><p>In cognitive science and psycholinguistics there is kind of analogue debate to explain the formation of concepts meaning. A variety of frameworks such embodiment, symbolism or representationalism offer complementary answers to how concepts and meaning is grounded by cognitive systems. Among the open fronts, an interesting topic is the explanation of <strong>abstract <em>vs</em> concrete meaning</strong>: how can abstract concepts be explained in terms of mental representations?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/356/0*mIS1FZg_o5pELhTa" /><figcaption>An illustration showing the abstract concept grounding problem. Source: <a href="https://royalsocietypublishing.org/doi/10.1098/rstb.2017.0121">https://royalsocietypublishing.org/doi/10.1098/rstb.2017.0121</a></figcaption></figure><h4>Linking both debates</h4><p>The debate about abstractness in language was one of the topics I felt more appealed to in my masters studies, so I thought it would be interesting to link this with the representationalism perspective from deep learning. I chose to replicate the methodology from Gurnee and Tegmark in a smaller scale, in order to investigate if abstract word meaning is represented within an LLM. To this end, there were several points to consider:</p><ul><li>First, how could I get data representing “the abstract meaning in language”? I employed for this question the data from a psycholinguistic study called <strong>the Glasgow norms</strong> (Scott et al., 2019), where 829 participants were surveyed to rate <strong>9 psycholinguistic dimensions</strong> related to a corpora of 5553 English words. The dimensions were <em>Arousal, Valence, Dominance, Concreteness, Imageability, Familiarity, Age of Acquisition, Semantic size </em>and<em> Gender association</em>. For my study, only the first seven dimensions and non-polysemic nouns –2440 words– were taken into consideration.</li><li>As a methodology, I followed a similar approach to Gurnee and Tegmark’s paper by predicting the Glasgow Norms scores sorely from the LLM model internal activations using probes. This question is very interesting because the source of the Glasgow Norms ratings is subjective data provided by many participants in a psycholinguistic study, but still aimed to represent the general population. The central question of my research is if these global subjective measures from the Glasgow Norm study could be also encoded on such a different information source like LLM hidden activations?</li><li>I had to choose a LLM model, and back when I started my master thesis, <a href="https://mistral.ai/news/announcing-mistral-7b"><strong>Mistral-7b</strong></a> was a good candidate: an open-source light LLM with 32 layers and 4096 hidden units (artificial neurons) at the end of each transformer block.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/262/1*Zu5LqYOaBhpTjZrsdgIoLA.png" /><figcaption><em>Model architecture hyperparameters for Mistral-7B (Jiang et al., 2023)</em></figcaption></figure><h4><strong>Hypothesis and Experimental Design</strong></h4><p>As with any investigation, I started with several hypotheses.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rO18JERXCzLvhdyOydkVJA.png" /><figcaption>Hypothesis followed in my study</figcaption></figure><p>To answer some of these questions, a total of 9 experiment conditions associated with 9 different prompt templates were designed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nZsCi7_kMq2TMFMGhbjCVw.png" /><figcaption><em>Experiments conditions, given by different prompt templates where the words were inserted</em></figcaption></figure><p>The selected words from the Glasgow Norms studies were combined and inserted into the 9 prompt templates. Mistral-7b was prompted at temperature zero with each of these prompts, and all the data with the LLM outputs and the hidden neurons activities was recorded for later analysis with the probes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*b_8kxL_knUjn3WZc" /><figcaption>Experimental Design on my work</figcaption></figure><h4>Data Preparation</h4><h4><strong>&gt; Internal LLM Activation Data</strong></h4><p>How to decode the activations from a large language model? Gurmee and Tenmark experimented with several LLMs of different sizes, some of them as big as Llama-70b. For my work I used <a href="https://medium.com/the-dl/how-to-use-pytorch-hooks-5041d777f904"><strong>pytorch hooks </strong></a>to analyze the 32 output-block layers in Mistral-7b (the final output of each transformer block layer). I selected each of these layers within the transformer for simplification and under the assumption that the richest representations could be found in those output layers of each transformer block.</p><p>Checking Mistral-7b responses to the combination of words with experimental conditions prompts was one of the funniest parts of my work. You can check Mistral-7b <a href="https://www.kaggle.com/datasets/fhipol/all-mistral-prompts-to-concepts">responses in this dataset</a>. Mistral is quite a small model, so the responses are gibberish or repetitive in some cases, but you will probably already notice some patterns there.</p><p>The activations were recorded for the first 50 feedforward passes within the LLM, and for six hidden layers (<em>1th</em>, <em>4th</em>, <em>11th</em>, <em>18th</em>, <em>25th</em> and <em>32th</em>). Thus, having 2240 words, for each of the experiments a total of 732000 embeddings were recorded.</p><p><em>2440 words </em>×<em> 50 feed-forward passes </em>×<em> 6 sampled layers = 732000 embeddings per prompt template</em></p><p>So in total, a dataset of 6048000 embeddings (activations vectors) recorded all the activations for all the combination of words, templates, hidden layers and feed-forward passes. All this embedding activation data was joined with the associated word-ratings from the Glasgow Norms. You also can check the whole final activation dataset <a href="https://www.kaggle.com/datasets/fhipol/dataframe-mistral-activations-for-glasgow-norms">here in Kaggle</a>.</p><p><strong>&gt; Predicting Glasgow Dimensions using Probes</strong></p><p>Having cooked all the data from the Mistral-7b activations for all the prompts, it is time to check if the Glasgow Norms data scores were encoded somehow on the Mistral-7b. I recorded the transformer output-block layer activations for each combination of template, <em>nth_layer</em> and psychological dimension. Thus, having 8 templates, 6 sampled layers and 9 psychological dimensions, 432 different probes were trained. The hidden activations were reduced using Principal Component Analysis (PCA), scaled and normalized layer-wise, and then used to train all the linear-regression probes used to predict the Glasgow Norms psychological dimensions.</p><p>In their work, Gurmee and Tenmark indeed tested the linear hypothesis via different kinds of probes, from linear one and non-linear probes to test if the linear hypothesis holds.</p><iframe src="https://cdn.embedly.com/widgets/media.html?type=text%2Fhtml&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;schema=twitter&amp;url=https%3A//x.com/wesg52/status/1709551572399907061&amp;image=" width="500" height="281" frameborder="0" scrolling="no"><a href="https://medium.com/media/5fc20de6cc884cae253a44ae47430c47/href">https://medium.com/media/5fc20de6cc884cae253a44ae47430c47/href</a></iframe><p>However in my study, for simplification only linear regression probes with Lasso regularization were used. These probes were fitted to predict the Mistral-7b model activations to the Glasgow Norms scores. You can review the notebooks and <a href="https://github.com/fhipol/cogsci-llm-concepts/tree/master/cogsci/notebooks">code in this repo</a>.</p><h4>Results</h4><p>The performance of the probes to predict the Glasgow scores (psychological dimensions) based on the LLM activations was measured using R² scores. Each probe on each layer is a linear regression model taking as input the activations for that layer during inference, and predicting as response the associated Glasgow Norm scores for a given psychological dimension and word.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-YMMkhKIEWTfDUILUiNNeQ.png" /><figcaption>Example of <em>Probes predictions (black dots) to Valence psychological dimension given in Glasgow Norm (orange line) for experiment-4th (</em><strong><em>what does {word} make you feel</em></strong><em>), for layers 1th, 17th and 32th. In general, the linear probes on deeper layers predict more closely the given Glasgow Norm dimension, pointing out to better encoding of the dimensions on the deeper model activation.</em></figcaption></figure><p>You can check all the Probe R² results in the following diagram, where each graph corresponds to a prompt template (1th to 9th); each colored series is a Glasgow Norm psychological dimension, the x-axis represents hidden layers, and the y-axis each Probe R² performance:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*E8JKD95gZUlX3UBu" /></figure><p>The most insightful results from the study can be seen in these graphs:</p><ul><li><strong>Probe performance</strong>, measured by R² scores, generally improved quite significantly in the deeper layers of the model, suggesting that the information relevant to psychological dimensions is better represented deeper in the model.</li><li><strong>Concreteness and Imageability</strong> were the psychological dimensions for which probes performed best across most experimental conditions. Concreteness reached an R² score of up to 0.7 in deeper layers for the second experiment (<em>“what does {word} mean?”</em>). Imageability, being highly correlated with Concreteness, was the second-best predicted dimension.</li><li><strong>Prompts providing context</strong>, like <em>“what does {word} mean?”</em> (Experiment 1th), resulted in better R² scores across most dimensions compared to the neutral condition where only the word was provided (Experiment 2th). The neutral condition yielded very low probe scores, indicating the importance of context.</li><li>Some specific prompts influenced the encoding of related dimensions; for example, the <em>“What does {word} make you feel?”</em> prompt (Experiment 4th) showed improved layer-to-layer performance for the Valence dimension.</li><li>I find this result one of the most impressive in my study: In contrast to the other experimental conditions where Concreteness and Imageability were typically the best-predicted dimensions by the linear probes, <strong>the <em>“spell the {word}”</em> prompt showed a different R² probes hierarchy</strong>. Specifically, the <strong>Age-of-Acquisition dimension </strong>was almost the best predicted dimension by the probes across layers in this condition (but for other templates it wasn’t represented so much!). Note that In the Glasgow Norms Study, Age of Acquisition (AoA) is a psycholinguistic dimension where participants estimated the typical age at which they learned a given word. This was one of the prompt templates differing more from the rest — as it didn’t involve “thinking” about the meaning of the word, but just spelling of it –, and somehow on Mistral-7b internal activations, Age-of-Acquisition estimations was encoded for this task! Soon to speculate, but could be the LLM encoding somehow the link to a word’s spelling complexity and the age at which children typically master it?</li><li>The success of linear probes in capturing significant variance in psychological dimensions supports the linear representation encoding hypothesis, suggesting these features are encoded as linear combinations of neuron activation within the model.</li></ul><h4>Discussion</h4><p>In this work it was discussed how Mistral-7b appears encode information related to abstractness, and this tends to show up deeper inside the model’s layers. Depending on how the model was prompted about a word, the internal patterns changed significantly, indicating that these representations are tied to the task or context given to the model.</p><p>I generally find fascinating that the ratings from a psycholinguistic study could be partially encoded on the internal activations of a simple LLM like Mistral-7b. Of course more extensive research is needed in this direction, but there could be also practical applications: for example, could we measure the “<em>abstractness</em>”, “<em>ambiguity</em>” or even psycholinguistic variables of a LLM during inference? these representations could be used as signals to make inference better. Imagine a model that, when its internal “hesitance signal” leaps, requests for more details instead of confidently guessing, helping to reduce those weird AI “hallucinations”. A new frontier could be open, and further research in LLMs interpretability and representationism would bring new insights.</p><h4><strong>Bibliography</strong></h4><ul><li>Gurnee, W., &amp; Tegmark, M. (2023). Language models represent space and time. <em>arXiv preprint arXiv:2310.02207</em>.</li><li>Scott, G. G., Keitel, A., Becirspahic, M., Yao, B., &amp; Sereno, S. C. (2019). The Glasgow Norms: Ratings of 5,500 words on nine scales. Behavior research methods, 51, 1258–1270.</li><li>Borghi Anna M., Barca Laura, Binkofski Ferdinand and Tummolini Luca 2018 Varieties of abstract concepts: development, use and representation in the brainPhil. Trans. R. Soc. B37320170121</li><li>Rączaszek-Leonardi, J., Nomikou, I., Rohlfing, K. J., &amp; Deacon, T. W. (2018). Language development from an ecological perspective: Ecologically valid ways to abstract symbols. Ecological Psychology, 30(1), 39–73. <a href="https://doi.org/10.1080/10407413.2017.1410387">https://doi.org/10.1080/10407413.2017.1410387</a></li><li>Tishby, N., Pereira, F. C., &amp; Bialek, W. (2000). The information bottleneck method. <em>arXiv preprint physics/0004057</em>.</li><li>Elhage, N., Hume, T., Olsson, C., Schiefer, N., Henighan, T., Kravec, S., … &amp; Olah, C. (2022). Toy models of superposition. <em>arXiv preprint arXiv:2209.10652</em>.</li><li>Kousta, S. T., Vigliocco, G., Vinson, D. P., Andrews, M., &amp; Del Campo, E. (2011). The representation of abstract words: why emotion matters. Journal of Experimental Psychology: General, 140(1), 14.</li><li>Jiang, A. Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D. S., Casas, D. D. L., … &amp; Sayed, W. E. (2023). Mistral 7B. <em>arXiv preprint arXiv:2310.06825</em></li></ul><p>You can check the Notebooks and the code used <a href="https://github.com/fhipol/cogsci-llm-concepts">in this Github Repo</a>, <a href="https://www.kaggle.com/fhipol/datasets">the datasets on Kaggle</a>, or <a href="https://zenodo.org/records/15812557">my full master thesis here</a>. I also want to thank Julian Zubek for his support as a mentor and supervision of my Master’s thesis.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9bd8038b6a7e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Meaning and Embodiment Theories in Psycholinguistics]]></title>
            <link>https://medium.com/@fhipol/meaning-and-embodiment-theories-in-psycholinguistics-2b6c3cc077ad?source=rss-29cd67b1dbf8------2</link>
            <guid isPermaLink="false">https://medium.com/p/2b6c3cc077ad</guid>
            <category><![CDATA[psycholinguistics]]></category>
            <category><![CDATA[cognitive-science]]></category>
            <category><![CDATA[meaning]]></category>
            <category><![CDATA[embodiment]]></category>
            <dc:creator><![CDATA[Fernando Jesus Garcia Hipola]]></dc:creator>
            <pubDate>Sat, 05 Apr 2025 19:09:53 GMT</pubDate>
            <atom:updated>2025-04-05T19:09:53.740Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wnSaASiTC70mLNzAJzotAQ.jpeg" /><figcaption>Photo from <a href="https://pixabay.com/photos/sound-space-network-wires-wire-3851251/">Pixabay</a></figcaption></figure><p>Meaning is a general topic that has brought the interest of diverse disciplines from a long time ago. Philosophy and linguistics were the first ones that addressed this problem:<em> </em>should we consider meaning as a combination of all the features that a word refers to, like the ingredients of a cake? Or is it more a problem about semiotics and the relationship between a symbol and the concept it designates? Is this concept indeed something objective which exists in the real world, or more a subjective construction from the subjects who create them? Many of these questions still do not have clear answers and depend on how the problem is formulated, or how the boundaries in the definition of meaning are established. As a contemporary science, psycholinguistics has been highly influenced by these questions from older disciplines, but now its goal pursues an objective truth based on empirical methods rather than general and abstract debates. Nevertheless, it is worth mentioning the importance of two contributions from philosophy that in my opinion highly impacted the direction in which the topic of meaning has been addressed by the discipline.</p><p>The first contribution is the notion of Wittgenstein about <strong>family resemblances</strong> (Wittgenstein, 1953). According to him, it is almost impossible to find a distinctive common feature to all the members of a category. The instances are related between them based on similarities which are not necessarily shared by all, like the members in a family do. Also, it is well known the analogy with games, where Wittgenstein highlights how difficult it is to find a clearly definitory unique feature for all the elements involved. In this way, dictionary-like descriptions cannot be seen as a sufficient method to define meanings, as we always might find exceptions to the simple defining rules.</p><p>The second contribution is from Searle which reacted against the prominence of symbolic theories about meaning and intentionality (Searle, 1980). He criticized some artificial intelligence promoters who supported that symbolic processing of information is sufficient to define the subjective capacity of cognition, this is the understanding and consciousness about meaning. His metaphor about a <strong>Chinese room</strong> is a recurrently quoted source to explain this problem: let us imagine a very good computer program which takes Chinese characters as inputs and delivers outputs to behave like a Chinese native speaker. This program indeed would not differ from a non-Chinese speaker who is inside a closed room simply taking inputs, and using a detailed instruction book about how process those inputs, can manipulate them and still deliver the correct outputs. In this second case we would not think that the person is indeed understanding the meaning of Chinese, as it is just executing instructions which give the correct output<em>.</em> This smart analogy serves to Searle as an argument that symbol manipulation must be <strong>grounded</strong> in the same kind of inner mechanism than human cognitive systems to be considered intentional and rich in meaning: otherwise, they are just blind systems which do not create any kind of correlated experiences.</p><p>Seale’s view undermined pure symbolic theories of cognition, but still structural models which rely on statistical co-occurrence in words are relevant in the field. They are useful for successfully modeling and predicting many aspects in experiments related to semantics like word retrieval or priming (Burgess &amp; Lund, 1997). Some of these models like <em>LSA</em> (Landauer et at, 1990) are based on the occurrence of words in large amounts of text (corpora), and they define the meaning as a function of the semantic distance between the words computed by these methods. Semantic networks are also useful models to explain the organization of words in terms of nodes and relationships between them.</p><p>While all these frameworks do not solve the symbol grounding problem, they show statistical patterns between symbols that capture a lot of information which is needed to define meaning: despite grounding on external elements is necessary to account for intentionality, statistics pattern might still play a relevant role in defining what meaning is.</p><h4>Embodiment Theories in Psycholinguistics</h4><p>The limitations of pure symbolism and the necessity to find a grounding solution led us to the current theories in psycholinguistics. Pure symbolic theories have been confronted by empirical data supporting embodied representations: for example, fMRI shows correlated activation between language words processing and modular areas activation which better represent the meaning of these words. All the modern theories include in its formulations some degree of embodiment, but they differ in the extent in which it is crucial for meaning realization, the modality of representations and how the information is processed. I will take an overview of the current classification by Meteyard et al (2012), where he differentiates between strong, weak, and secondary embodiment.</p><p><strong>Strong embodiment theories</strong> claim that embodiment experiences and representations are a necessary condition for meaning, and concepts rely entirely on this kind of phenomena. Embodiment is in this way “the material” which constitutes meaning, rather than being a secondary product. When I represent a concept like <em>elephant</em>, I rely on diverse perceptual representations that are dependent on many different modules: the nature of the concept being represented and how it is experienced by the subject will determine the predominant modalities of representation. For example, in the case of an animal concept, the visual perception features or the sound it makes could be prevalent, but in the case of a tool, the functional and motor experience would have a major weight on its representation. Strong embodiment theories are multi-modal, highly distributed and related to simulation theories (the comprehension of concepts and meaning requires full simulation by the subject). As they do not differentiate between the nature of the concept itself and the embodied representation, they locate the structures responsible for meaning within the modules responsible for creating such embodied representations, namely sensory and motor systems. Among the theories we can highlight the index hypothesis by Glengerg and Roberton<strong> </strong>(2000), which on the concept of affordances and how they play a relevant role in the process of meaning constitution: the possibilities of interaction of a subject with the world and his own subjective experience are the direct constituents of what meaning is.</p><p><strong>Secondary embodiment theories</strong> reject that embodiment is the primary constituent of concepts. There are several perspectives among these theories, and thought they might concede that semantic representations are somehow correlated with embodied representation, they are neither constituted nor produced by them. The semantic system therefore interacts with the embodied system through by means of connectivity and interaction, but they are essentially independent from each other. We can mention Caramazza as one of the most relevant authors supporting these views: for him embodiment is a consequence of the spreading activation from the semantic system linked to the systems responsible for the embodied experiences (Mahon &amp; Caramazza, 2008). Other processes such as statistical pattern abstraction between the occurrence of symbols, logical processes or even categories led by biological evolution might explain the representation of some concepts without embodiment simulation. These theories agree that semantic representation relies ultimately on a structure independent from the rest of modules. This could be related to <strong>Converge Zones (CZs) i</strong>ntroduced by Damasio (1989) which represent abstract information taken from the embodied generating regions of the brain. As reviewed later, all embodiment theories relate from different perspectives to this idea of specific zones in the brain where the information from different modules converge. Damasio highlights that these zones are not where the subjective experience of the content happens — as damage in these regions do not affect the possibility of these experiences — , but the integration task which makes possible recognition. For secondary embodiment theories, the CZs play a role in reactivating the embodied representations following semantic abstraction and linking both mechanisms.</p><p><strong>Weak embodiment theories</strong> stand in an intermediary position by not considering embodiment as the constituent of meaning <em>per se</em>, but attributing a crucial role in conjunction within other processes. They predict that the brain areas responsible for constituting meaning are highly distributed through the brain, and although they are not the perceptual/motor areas as predicted by strong embodiment, they are located nearby those embodied experiential areas that support the processing task. Embodiment representations are therefore frequent constituents of concepts, but some degree of abstraction comes into play when creating meaning: it is not a pure simulation of motion and perceptual states, but just “a film” of the object represented. This is the position adopted by investigators like Pulvermüller who criticized the minimal embodiment theories from Caramazza and others. According to him minimal theories “<em>still fails to explain why action and perception systems exert causal effects on the processing of symbols from specific semantic classes</em>” (Pulvermüller, 2013).</p><p>In general, all the previous theories (weak, secondary, and strong embodiment) coexist because they are coherent with empirical evidence: each of them successfully explains certain phenomena and find limitations where the others succeed. Results from imaging data, lexico semantics, electrophysiological and other experiments can be taken to support one or other theories depending on the interpretation of the results as pointed by Vigliocco et al (2012). These investigators think that all the theories agree in some form of converge zones which make possible the interaction between embodied representations and meaning.</p><p>Some researchers like Kyle Simmons and Lawrence W. Barsalou (2003) thinks that all these theories are not entirely mutually exclusive. They proposed the <strong>Conceptual Topography Theory</strong> which further elaborates Damasio’s converge zones (CZs) by differentiating some kind of CZs with different roles in the meaning creation process:</p><ul><li>Analytic CZs would be responsible for abstracting simple properties from features within modules (namely the smoothness, edges patterns in visual modules or the depth or tones in sound modules).</li><li>Holistic CZs would combine these properties and others to find higher level patterns (for example how the position of the eyes and the nose define a face).</li><li>Modality CZs are more specific to statistical patterns, correlations and abstraction operations over the features and the properties to making then available for other modules.</li><li>Intermodular CZs act as a liaison between modules integrating different kinds of representations, gluing them into a unitary representation which can be used to reactivate the initial features later on.</li></ul><p>To explain some limitations of CZs they introduce the SIT principle by which similar representations should be localized topologically nearby on the CZs. The spread of these representations depends on the concept nature: for example, animals would be represented closer to each other as they highly depend on common visual patterns. However, tools would be more spread between them — topologically adjacent but further away in terms of distance — as the properties and functions they rely on are more diverse. This could explain why some impairments affect the retrieval of animal categories rather than tools, as they are more localized in a specific region dependent on visual stimuli and therefore susceptible to loss of functionality in case of damage.</p><h4>Embodiment and abstraction</h4><p>A common limitation from embodied theories is explaining obscure concepts whose representations do not depend on concrete experience. Abstract concepts like love or justice do not rely on specific embodied content. A reply to this objection is that these kinds of concepts are more based on other types of embodying like internal states or emotions rather than perceptual phenomena. An explanation from less embodied theories would probably claim that these concepts are a product of a more extensive abstraction in high level processing of the information. I think abstraction is indeed the most interesting topic which should be addressed, or how stop-words which are considered meaningless (namely prepositions, conjunctions) could be explained by the embodiment theories. Are abstract meanings just another modality of embodiment more difficult to identify, or is there in fact another independent mechanism responsible for them?</p><h4><strong>Bibliography</strong></h4><ul><li>Traxler, M. J. (2012). Introduction to psycholinguistics: Understanding language science. Chichester, West Sussex: WileyBlackwell. (p.89–97)</li><li>Searle, J. R. (1980). Minds, brains, and programs. Behavioral and brain sciences, 3(3), 417–424.</li><li>Damasio, A. R. (1989). Time-locked multiregional retroactivation: A systems-level proposal for the neural substrates of recall and recognition. Cognition, 33(1–2), 25–62.</li><li>Wittgenstein, L. (2010). Philosophical investigations. John Wiley &amp; Sons.</li><li>Meteyard, L., Cuadrado, S. R., Bahrami, B., &amp; Vigliocco, G. (2012). Coming of age: A review of embodiment and the neuroscience of semantics. Cortex, 48(7), 788–804.</li><li>Deerwester, S., Dumais, S. T., Furnas, G. W., Landauer, T. K., &amp; Harshman, R. (1990). Indexing by latent semantic analysis. Journal of the American society for information science, 41(6), 391–407.</li><li>Glenberg, A. M., &amp; Robertson, D. A. (2000). Symbol grounding and meaning: A comparison of high-dimensional and embodied theories of meaning. Journal of memory and language, 43(3), 379–401.</li><li>Pulvermüller, F. (2013). Semantic embodiment, disembodiment of disembodiment? In search of meaning in modules and neuron circuits. Brain and language, 127(1), 86–103.</li><li>Simmons, W. K., &amp; Barsalou, L. W. (2003). The similarity-in-topography principle: Reconciling theories of conceptual deficits. Cognitive neuropsychology, 20(3–6), 451–486.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2b6c3cc077ad" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Blockchain will change the game]]></title>
            <link>https://medium.com/@fhipol/how-blockchain-will-change-the-game-f8267c6bf485?source=rss-29cd67b1dbf8------2</link>
            <guid isPermaLink="false">https://medium.com/p/f8267c6bf485</guid>
            <category><![CDATA[blockchain]]></category>
            <dc:creator><![CDATA[Fernando Jesus Garcia Hipola]]></dc:creator>
            <pubDate>Wed, 19 Sep 2018 11:41:47 GMT</pubDate>
            <atom:updated>2018-09-19T11:41:47.804Z</atom:updated>
            <content:encoded><![CDATA[<p>Blockchain here, blockchain there. Everyone is talking about it, and it seems to be a magical word in every business plan or start-up claim. Hard to distinguish if it is just a fashion word, or if it is really the promising revolutionary technology which will change the world in the coming years. When the top companies, professional services networks, tech experts or investors are preaching its revolutionary potential, probably it is somehow true that blockchain will have a determinant role in the upcoming new digital revolution. But also, these voices are mixed with those of gurus, opportunity seekers and marketers which just want to make noise and that their brand appears associated with it.</p><p>I will admit that the big success of Bitcoin and ICOS in the last years attracted my attention in blockchain: rising numbers, big headlines in newspapers, friends talking about investing on Bitcoin by relying more in the lucky success of an acquaintance than in a deep reflexive analysis. All these were factors which leaded me to read a little more about blockchain: first watching high-level explanation of its intricate technology, later reading articles or those dense white and yellow papers of Ethereum. Hard to pick the details for a non-specialist like me, but enough for understanding the clever key features which make possible the technology. And this is the first big challenge of blockchain: how to communicate efficiently to the masses what blockchain is, what it is for and why it is worth. This last question is which I really wanted to answer.</p><p>My first reflections about that question began when I registered in Cryptokitties, a Tamagotchi-like game which runs in Ethereum, but it is accessible via web using a browser component. I enjoyed spending time on it, but I hardly found any interest for the average user, in whom the true potential blockchain success would ultimately rely. This game and other blockchain applications are already existing business models or services in which decentralization has been added as a patch: I do not undervalue it, but I think the manner how it is being regarded is not the only reason why blockchain will be revolutionary. This is the impression I get: right now, most of the blockchain potential and promotion is being focused on values such a decentralization, privacy and independence, which are only appealing for concrete niches in the market: digital geeks, tech revolutionaries and ideologists, people who dreams with a system decentralized from institutions, and people who really cares about its private data (I can assure people don´t generally care so much about their data as it seems). Blockchain as a service is great for all these niches, but it not the disruptive feature which will prompt its high scale success. Blockchain marketing and popularity in the last years has highly increased the interest on it, and consequently the niches I have talked about have grown. But this interest and marketing fame is only the air which inflates a big bubble: when it stops, the bubble deflates, or even it explodes in the worst case. Blockchain true success won’t come from selling its values to people who firstly is not interested on it, but in really making a real true disruptive change on how the world works.</p><p>To find this true disruptive value of blockchain we may consider its similarity with internet as potential disruptive technology. Internet started as dreamed project with many benefits: in the early ages it was designed as an efficient system for military and scientific applications, and later for commercial and society communication. It attracted the attention of many people who liked the ideologic potential on it: free knowledge share, collaboration, end of physical borders… all these were useful and relevant qualities which made internet promising, but not still disruptive. This is the current stage of blockchain. What really made internet disruptive was the same reason why paper newspapers, physical retailers, media and so many other traditional businesses died or had to quickly adapt to the new digital environment it created: <strong>a much more efficient system</strong> <strong>which had a direct impact in costs, and therefore in final product or services prices.</strong> And that´s what really blockchain will repeat and prompt its true key disrupt success.</p><p>Internet removed the necessity of physically going in person to a shop to buy a product: not all physical shops have disappeared, but only those which aggregates the value that internet cannot provide have survived. So many costs were removed by internet, and so many will be by blockchain. However, it stills lacks from a medium thought which people really can cope and interact with it intuitively. In fact, internet didn’t rely in only one medium to do it: from the raw internet (lines of code and technical stuff), there were more and more increasing interfaces which helped users to better understand and interact with it: first there were graphical systems and cunning metaphors as navigators, then virtual communities, someone-to-ask as browsers, user profiles in the social networks, and later total access availability by any device through the cloud. All these effective increasing interfaces is what blockchain lacks right now, but not for long. What I am saying is nothing new, and many companies are working in bringing the blockchain benefits through intuitive services and interfaces. It´s a matter of time that one of them gets the key to penetrate massively in the market: for sure it will imply a strong impact in how supply chains, services, businesses and in general the society.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f8267c6bf485" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>