Ram Bharadwaj

Exploring the Platonic Representation Hypothesis Beyond In-Distribution Data

The Platonic Representation Hypothesis (PRH) suggests that models trained with different objectives and on various modalities can converge to a shared statistical understanding of reality. While this is an intriguing idea, initial experiments in the paper focused on image-based models (like ViT) trained on the same pretraining (ImageNet) dataset. This raises an important question: Does PRH hold only when models are trained on data from the same distribution? To explore this question, the experiment was extended to ImageNet-O—a dataset specifically designed with out-of-distribution (OOD) images compared to ImageNet. Using ImageNet-O, the correlation analysis of alignment scores across various metrics for image classification models was re-evaluated. The outcome? PRH holds true in the OOD setting as well, which challenges the notion that a shared data distribution is a prerequisite for this convergence. This observation carries significant implications for AI alignment research, suggesting that a deeper underlying structure may govern how models develop representations of reality, even when the training data differs. Below are the correlation plots comparing results from the original in-distribution experiment to those with OOD data: But does this mean that the model's align even for purely randomly generated data? The answer is NO. This plot shows the correlation of the alignment scores for the models on purely randomly generated images. One key observation here is: * On randomly generated data, PRH is not so true, I.e, the models did not share a statistical model of reality. * On OOD data of ImageNet-O dataset they have shared, and have high correlation. But on ImageNet-O the predictions are wrong but with higher confidence.I.e, all the models classified the images wrongly on this outlier data yet predictably. This is particularly interesting since we can infer that they fail predictably. The notebook documenting these experiments is available h

13Oct 20, 2024

Ram Bharadwaj

Message

Scaling Laws for LLM Based Data Compression

Introduction The scaling laws for neural language models showed that cross-entropy loss follows a power law in three factors: 1. Dataset size 2. Model parameters 3. Training compute (steps or epochs) The lower the cross-entropy loss, the better the model’s next-token prediction. Since prediction and compression are deeply linked (The...

Aug 5, 20258

Exploring the Platonic Representation Hypothesis Beyond In-Distribution Data

Oct 20, 202413

Understanding Hidden Computations in Chain-of-Thought Reasoning

Recent work has demonstrated that transformer models can perform complex reasoning tasks using Chain-of-Thought (COT) prompting, even when the COT is replaced with filler characters. This post summarizes our investigation into methods for decoding these hidden computations, focusing on the 3SUM task. Background 1. **Chain-of-Thought (COT) Prompting**: A technique that...

Aug 24, 20246

Goal-misgeneralization is ELK-hard

Consider an adversarial training-scheme for solving goal-misgeneralization, ( here i consider Redwood Research’s work on “Adversarial Training for High-Stakes Reliability”). Consider a model that was trained to perform a specific task. To guarantee worst-case performance for this model, we need to have bounds for its outputs in adversarial examples. However,...

Jun 10, 20232

Hutter-Prize for Prompts

The aim of the Hutter Prize is to compress the first 1GB of Wikipedia to the smallest possible size. From the AIXI standpoint, compression is equal to AI, and if we can compress this to the ideal size (75MB according to Shannon's lower estimate), then the compression algorithm is equivalent...

Mar 24, 20235

The AGI needs to be honest

Imagine that you are a trained mathematician and you have been assigned the job of testing an arbitrarily intelligent chatbot for its intelligence. You being knowledgeable about a fair amount of computer-science theory won’t test it with the likes of Turing-test or similar, since such a bot might not have...

Oct 16, 20212

LESSWRONG
LW

LESSWRONG
LW

Ram Bharadwaj

Ram Bharadwaj

Ram Bharadwaj

Exploring the Platonic Representation Hypothesis Beyond In-Distribution Data

Scaling Laws for LLM Based Data Compression

Understanding Hidden Computations in Chain-of-Thought Reasoning

Hutter-Prize for Prompts

Ram Bharadwaj

Scaling Laws for LLM Based Data Compression

Exploring the Platonic Representation Hypothesis Beyond In-Distribution Data

Understanding Hidden Computations in Chain-of-Thought Reasoning

Goal-misgeneralization is ELK-hard

Hutter-Prize for Prompts

The AGI needs to be honest

Exploring the Platonic Representation Hypothesis Beyond In-Distribution Data

Scaling Laws for LLM Based Data Compression

Understanding Hidden Computations in Chain-of-Thought Reasoning

Hutter-Prize for Prompts

Scaling Laws for LLM Based Data Compression

Exploring the Platonic Representation Hypothesis Beyond In-Distribution Data

Understanding Hidden Computations in Chain-of-Thought Reasoning

Goal-misgeneralization is ELK-hard

Hutter-Prize for Prompts

The AGI needs to be honest