<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>machine learning on Suriya's site</title><link>https://subygan.com/tech/ML/</link><description>Recent content in machine learning on Suriya's site</description><generator>Hugo</generator><language>en</language><copyright>This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.</copyright><item><title>RL, in pictures and videos</title><link>https://subygan.com/tech/ML/rl_in_pictures/</link><pubDate>Sat, 25 Apr 2026 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/rl_in_pictures/</guid><description>&lt;h2 id="drone-trained-by-rl-beat-the-worlds-best-human-pilots-2023">drone trained by RL beat the world&amp;rsquo;s best human pilots (2023)&lt;/h2>
&lt;p>UZH, 2023. Trained in simulation. Onboard camera only. Beat three FPV world champions on a real track.&lt;/p></description></item><item><title>practical deep learning</title><link>https://subygan.com/tech/ML/practical_deep_learning/</link><pubDate>Sat, 15 Feb 2025 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/practical_deep_learning/</guid><description>&lt;p>notes from &lt;a href="https://www.fast.ai/">fast.ai&lt;/a>
 Practical Deep Learning course.&lt;/p></description></item><item><title>photogrammetry/videogrammetry or the dark art of capturing 2d and generating 3d models</title><link>https://subygan.com/tech/ML/photogrammetry/</link><pubDate>Fri, 11 Oct 2024 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/photogrammetry/</guid><description>&lt;h1 id="capturing">capturing&lt;/h1>
&lt;h2 id="questions">questions:&lt;/h2>
&lt;p>Is it possible to go from, photo/video to high fidelity 3d model, with strong support for texture, without losing a lot of details on noise?&lt;/p></description></item><item><title>ml journal</title><link>https://subygan.com/tech/ML/ML_diary/</link><pubDate>Thu, 08 Aug 2024 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/ML_diary/</guid><description>&lt;h1 id="092723">09.27.23&lt;/h1>
&lt;ul>
&lt;li>Tried ML model for face classification with,&lt;/li>
&lt;/ul>
&lt;div class="highlight">&lt;pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;">&lt;code class="language-python" data-lang="python">&lt;span style="display:flex;">&lt;span>
&lt;/span>&lt;/span>&lt;span style="display:flex;">&lt;span> self&lt;span style="color:#f92672">.&lt;/span>backbone &lt;span style="color:#f92672">=&lt;/span> torch&lt;span style="color:#f92672">.&lt;/span>nn&lt;span style="color:#f92672">.&lt;/span>Sequential(
&lt;/span>&lt;/span>&lt;span style="display:flex;">&lt;span> torch&lt;span style="color:#f92672">.&lt;/span>nn&lt;span style="color:#f92672">.&lt;/span>Conv2d(&lt;span style="color:#ae81ff">3&lt;/span>,&lt;span style="color:#ae81ff">64&lt;/span>,kernel_size&lt;span style="color:#f92672">=&lt;/span>&lt;span style="color:#ae81ff">7&lt;/span>,stride&lt;span style="color:#f92672">=&lt;/span>&lt;span style="color:#ae81ff">4&lt;/span>),
&lt;/span>&lt;/span>&lt;span style="display:flex;">&lt;span> torch&lt;span style="color:#f92672">.&lt;/span>nn&lt;span style="color:#f92672">.&lt;/span>Conv2d(&lt;span style="color:#ae81ff">64&lt;/span>,&lt;span style="color:#ae81ff">128&lt;/span>, kernel_size&lt;span style="color:#f92672">=&lt;/span>&lt;span style="color:#ae81ff">3&lt;/span>, stride&lt;span style="color:#f92672">=&lt;/span>&lt;span style="color:#ae81ff">2&lt;/span>),
&lt;/span>&lt;/span>&lt;span style="display:flex;">&lt;span> torch&lt;span style="color:#f92672">.&lt;/span>nn&lt;span style="color:#f92672">.&lt;/span>Conv2d(&lt;span style="color:#ae81ff">128&lt;/span>,&lt;span style="color:#ae81ff">256&lt;/span>,kernel_size&lt;span style="color:#f92672">=&lt;/span>&lt;span style="color:#ae81ff">3&lt;/span>, stride&lt;span style="color:#f92672">=&lt;/span>&lt;span style="color:#ae81ff">2&lt;/span>),
&lt;/span>&lt;/span>&lt;span style="display:flex;">&lt;span> torch&lt;span style="color:#f92672">.&lt;/span>nn&lt;span style="color:#f92672">.&lt;/span>Conv2d(&lt;span style="color:#ae81ff">256&lt;/span>, &lt;span style="color:#ae81ff">512&lt;/span>,kernel_size&lt;span style="color:#f92672">=&lt;/span>&lt;span style="color:#ae81ff">3&lt;/span>, stride&lt;span style="color:#f92672">=&lt;/span>&lt;span style="color:#ae81ff">2&lt;/span>),
&lt;/span>&lt;/span>&lt;span style="display:flex;">&lt;span> torch&lt;span style="color:#f92672">.&lt;/span>nn&lt;span style="color:#f92672">.&lt;/span>Conv2d(&lt;span style="color:#ae81ff">512&lt;/span>, &lt;span style="color:#ae81ff">1024&lt;/span>,kernel_size&lt;span style="color:#f92672">=&lt;/span>&lt;span style="color:#ae81ff">3&lt;/span>, stride&lt;span style="color:#f92672">=&lt;/span>&lt;span style="color:#ae81ff">2&lt;/span>),
&lt;/span>&lt;/span>&lt;span style="display:flex;">&lt;span> )
&lt;/span>&lt;/span>&lt;span style="display:flex;">&lt;span>
&lt;/span>&lt;/span>&lt;span style="display:flex;">&lt;span> self&lt;span style="color:#f92672">.&lt;/span>cls_layer &lt;span style="color:#f92672">=&lt;/span> torch&lt;span style="color:#f92672">.&lt;/span>nn&lt;span style="color:#f92672">.&lt;/span>Sequential(
&lt;/span>&lt;/span>&lt;span style="display:flex;">&lt;span> torch&lt;span style="color:#f92672">.&lt;/span>nn&lt;span style="color:#f92672">.&lt;/span>Flatten(),
&lt;/span>&lt;/span>&lt;span style="display:flex;">&lt;span> torch&lt;span style="color:#f92672">.&lt;/span>nn&lt;span style="color:#f92672">.&lt;/span>Linear( &lt;span style="color:#ae81ff">4096&lt;/span>,num_classes),
&lt;/span>&lt;/span>&lt;span style="display:flex;">&lt;span> )
&lt;/span>&lt;/span>&lt;/code>&lt;/pre>&lt;/div>&lt;p>This model overfit spectacularly, training accuracy after 20 epochs was 95% validation accuracy was 15%. 🤯&lt;/p></description></item><item><title>word embedding models</title><link>https://subygan.com/tech/ML/word2vec/</link><pubDate>Sun, 04 Aug 2024 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/word2vec/</guid><description>&lt;p>&lt;strong>My aim for this blog was to keep it as not too technical as possible.&lt;/strong>&lt;/p>
&lt;p>Word embeddings is a way to represent words as a vector.
Which lets us do all kinds of trickery like how far a word like &amp;ldquo;King&amp;rdquo; is away from &amp;ldquo;Queen&amp;rdquo; and how they are related to each other with.
Very useful when building applications with semantic understanding.&lt;/p></description></item><item><title>streaming diarizer</title><link>https://subygan.com/tech/ML/streaming_diarizer/</link><pubDate>Sat, 20 Apr 2024 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/streaming_diarizer/</guid><description>&lt;p>&lt;a href="https://github.com/subygan/SpeechStream">repo&lt;/a>
&lt;/p>
&lt;ul>
&lt;li>&lt;a href="https://github.com/NVIDIA/NeMo/blob/main/tutorials/speaker_tasks/Speaker_Diarization_Training.ipynb">Nemo Speaker diarization training&lt;/a>
&lt;/li>
&lt;li>&lt;a href="https://github.com/NVIDIA/NeMo/blob/main/tutorials/tools/Multispeaker_Simulator.ipynb">MultiSpeaker Simulation&lt;/a>
&lt;/li>
&lt;/ul>
&lt;h3 id="data-sets">data sets&lt;/h3>
&lt;ul>
&lt;li>&lt;a href="https://www.openslr.org/12">LibriSpeech&lt;/a>
&lt;/li>
&lt;li>&lt;a href="https://github.com/joonson/voxconverse">VoxConverse&lt;/a>
&lt;/li>
&lt;/ul>
&lt;h2 id="resources">resources&lt;/h2>
&lt;ul>
&lt;li>&lt;a href="https://github.com/ufal/whisper_streaming">Whisper Streaming&lt;/a>
&lt;/li>
&lt;li>&lt;a href="https://ieeexplore.ieee.org/abstract/document/10446050">Diarist&lt;/a>
&lt;/li>
&lt;/ul></description></item><item><title>llm code generation notes</title><link>https://subygan.com/tech/ML/code_llm/</link><pubDate>Tue, 19 Mar 2024 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/code_llm/</guid><description>&lt;p>This is going to be binary analysis with LLM type thing
The idea is to see, if LLMs are able to understand the &amp;ldquo;language&amp;rdquo; of binaries and are able to retrieve the underlying vulnerability.&lt;/p></description></item><item><title>federated learning</title><link>https://subygan.com/tech/ML/federated_learning/</link><pubDate>Sun, 03 Mar 2024 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/federated_learning/</guid><description/></item><item><title>different types of graphs i come across</title><link>https://subygan.com/tech/ML/plots/</link><pubDate>Sun, 25 Feb 2024 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/plots/</guid><description>&lt;h3 id="pandas">pandas&lt;/h3></description></item><item><title>pandas fight</title><link>https://subygan.com/tech/ML/pandas/</link><pubDate>Sat, 24 Feb 2024 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/pandas/</guid><description>&lt;p>When a column has NaN it is shown as float value in a column.
need to use &lt;code>.fillna(&amp;quot;&amp;quot;)&lt;/code> to replace the &lt;code>Nan&lt;/code> with empty string.&lt;/p></description></item><item><title>graphs and llms</title><link>https://subygan.com/tech/ML/graph_llm/</link><pubDate>Sun, 11 Feb 2024 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/graph_llm/</guid><description>&lt;p>I&amp;rsquo;ll add the ones with the most fun:utilization first&lt;/p>
&lt;h3 id="building-graphs-using-llms">building graphs using llms:&lt;/h3>
&lt;p>This strategy comes clearly under the &lt;strong>by LLMs&lt;/strong> category and the most obvious one as well. Because of their semantic understanding, LLMs can be prompted into providing the triplets of relationship between the entities in a piece of text. Throw a piece of text, and it draws a graph of the relationship between subjects and objects
This is a relatively straightforward utilization of an LLM&amp;rsquo;s ability to find pattern in text.
eg. &lt;a href="https://news.ycombinator.com/item?id=34605772">GraphGPT&lt;/a>
w&lt;/p></description></item><item><title>research papers</title><link>https://subygan.com/tech/ML/papers/</link><pubDate>Fri, 09 Feb 2024 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/papers/</guid><description>&lt;ul>
&lt;li>
&lt;p>aiaac, AI incidents database.&lt;/p>
&lt;/li>
&lt;li>
&lt;p>&lt;a href="https://arxiv.org/pdf/2311.17035.pdf">Scalable Extraction of Traning Data from (Production) Language Models&lt;/a>
&lt;/p>
&lt;ul>
&lt;li>Making LLMs repeat the same word over and over again makes them emit the training data 3.5X better&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>
&lt;p>&lt;a href="//">The pain of Building copilots&lt;/a>
&lt;/p></description></item><item><title>ml guardrails</title><link>https://subygan.com/tech/ML/llm_guardrails/</link><pubDate>Tue, 06 Feb 2024 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/llm_guardrails/</guid><description>&lt;p>The base premise is that, aligning the model for the goal that it was created for is a long tail problem and requires multiple guardrails to control it&amp;rsquo;s output. Without this, the model could possibly go off the rails and create a lot of confusion.&lt;/p></description></item><item><title>prompt engineering</title><link>https://subygan.com/tech/ML/prompt_engineering/</link><pubDate>Tue, 23 Jan 2024 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/prompt_engineering/</guid><description>&lt;h2 id="in-context-learning">in context learning&lt;/h2>
&lt;h2 id="chain-of-thought">chain-of-thought&lt;/h2>
&lt;p>Broadly three classes of reasoning tasks
&amp;ldquo;let&amp;rsquo;s think step by step&amp;rdquo; seems to be the key phrase&lt;/p></description></item><item><title>purple llama</title><link>https://subygan.com/tech/ML/purple_llama/</link><pubDate>Mon, 11 Dec 2023 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/purple_llama/</guid><description>&lt;p>PurpleLlama is a new framework proposed by Meta to evaluate LLMs for their secure coding ability.&lt;/p>
&lt;h2 id="cyberseceval">cyberseceval&lt;/h2>
&lt;p>This is the core part of the research paper which has created a &lt;a href="https://github.com/facebookresearch/PurpleLlama/tree/main/CybersecurityBenchmarks">benchmark for LLMs&lt;/a>
 to test generated code.&lt;/p></description></item><item><title>membership inference attack</title><link>https://subygan.com/tech/ML/MIA/</link><pubDate>Sat, 09 Dec 2023 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/MIA/</guid><description>&lt;p>Membership inference Attack is a way to retrieve back the probability of the original training inputs.&lt;/p>
&lt;h3 id="notes">notes&lt;/h3>
&lt;ul>
&lt;li>&lt;a href="https://arxiv.org/pdf/1610.05820.pdf">https://arxiv.org/pdf/1610.05820.pdf&lt;/a>
&lt;/li>
&lt;/ul></description></item><item><title>intro to llm</title><link>https://subygan.com/tech/ML/Intro_to_llm/</link><pubDate>Thu, 23 Nov 2023 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/Intro_to_llm/</guid><description>&lt;p>My notes on Karpathy&amp;rsquo;s &lt;a href="https://youtu.be/zjkBMFhNj_g">Intro to LLM talk&lt;/a>
&lt;/p>
&lt;h3 id="llama-2-70b">llama-2-70b&lt;/h3>
&lt;ul>
&lt;li>comes with &lt;code>run.c&lt;/code> ~500 lines of c for it to be a self-contained runnable model.&lt;/li>
&lt;li>LLM are a lossy compression of the internet.&lt;/li>
&lt;li>In the process of trying to predict the next word, The model would have to understand the World that led to the creation of that next word.&lt;/li>
&lt;li>What the model does is closer to a dream, than determinism. The model only knows that something of a certain structure has to come there. It is not going to exactly replicate the world. It is dreaming about it and just like every dream the finer details could be hazy.&lt;/li>
&lt;li>Reversal curse, knowledge that can be accessed from one angle, cannot be accessed when from another angle. And it is hard to understand why, because even though the equations behind the models are well understood. It is hard to reason within the billions of parameters and figure out why the model came to a certain conclusion.&lt;/li>
&lt;/ul>
&lt;h3 id="assistant-training">assistant training&lt;/h3>
&lt;p>After training the Language model on a number of Epochs over the whole corpus of internet text.&lt;/p></description></item><item><title/><link>https://subygan.com/tech/ML/broadcasting_rules/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/broadcasting_rules/</guid><description/></item><item><title/><link>https://subygan.com/tech/ML/micrograd/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><guid>https://subygan.com/tech/ML/micrograd/</guid><description>&lt;!-- ---
emoji:
title: Micrograd
description: Notes from karpathy's micrograd
date: 
layout:
--- --></description></item></channel></rss>