Kirill Akhmetzyanov

AI of everything

[Anthropic @AnthropicAI]

In previous research, we found that Claude expresses over 3,000 values, like honesty and warmth. In new work, we asked how the values Claude expresses vary between Claude models and across languages.

We analyzed 300K+ anonymized conversations to find out.https://t.co/PgxsMXipt5

[Kirill Akhmetzyanov @Kirill_AI]

OpenAI Agent Builder was recently announced. It looks awesome! Has anyone tried it yet?

#openai #ai

[Kirill Akhmetzyanov @Kirill_AI]

But where to look for this specific problem?..

[Kirill Akhmetzyanov @Kirill_AI]

I can create the technical part. But how to create a product? No one taught me that

[Kirill Akhmetzyanov @Kirill_AI]

I understand that this requires solving a specific problem that clients are willing to pay for

[Kirill Akhmetzyanov @Kirill_AI]

Just a little confused with the amount of templates...

[Kirill Akhmetzyanov @Kirill_AI]

Ok, has anyone tried creating with n8n?

[Kyle Cranmer @KyleCranmer]

Tomorrow (Wednesday) I'll be giving a talk at the Mathematical and Scientific Machine Learning #MSML2021 conference. I'll talk about a physics-inspired problem that is not easy for ML and a toy version where we can compute the answer exactly.
https://t.co/ELKA9EOZte

[Kirill Akhmetzyanov @Kirill_AI]

Idea: automatic task creation for meta-learning based on UL.

NN creates embeddings. Clustering algorithm uniformly distributes cluster centers over the feature space and, based on the distance to the centers, creates pseudo labels. Meta-learning is run for the generated tasks

[Kirill Akhmetzyanov @Kirill_AI]

Today's review is about unsupervised learning + meta-learning!

🎓Unsupervised Learning via Meta-Learning
by Kyle Hsu (@kylehkhsu), Sergey Levine, Chelsea Finn

#shortreview #ML #AI #datascience #deeplearning #programming #machinelearning #100DaysOfCode

[Kirill Akhmetzyanov @Kirill_AI]

A gradient-based method for creating a task-dependent NAS is described.

Features for rapid learning:
1. Uses DARTS first order optimization.
2. Weights and architecture parameters are optimized simultaneously both in the inner loop and in the outer one.

[Kirill Akhmetzyanov @Kirill_AI]

🎓Towards Fast Adaptation of Neural Architectures with Meta Learning
by Dongze Lian, Yin Zheng, Yintao Xu, Yanxiong Lu et al.

#shortreview #ML #AI #datascience #deeplearning #programming #machinelearning #100DaysOfCode

[Kirill Akhmetzyanov @Kirill_AI]

M-NAS consists of several components:
1. Task-aware architecture generation - recurrent auto-encoder and LSTM controller.
2. Transferable knowledge modulation - a fully-connected layer and gate.
3. Efficient optimal weights estimation - inner and outer loop.

[Kirill Akhmetzyanov @Kirill_AI]

🎓M-NAS: Meta Neural Architecture Search
by Jiaxing Wang, Jiaxiang Wu, Haoli Bai, Jian Cheng

#shortreview #ML #AI #datascience #deeplearning #programming #machinelearning #100DaysOfCode

[Kirill Akhmetzyanov @Kirill_AI]

The method in the article combines optimization-based meta-learning and NAS.

The problem is splitted into two issues: 1) find an architecture controller that effectively generates models, 2) estimate optimal weights for generated model.

[Kirill Akhmetzyanov @Kirill_AI]

MetaNas combines two gradient methods: Reptile and DARTS.

Features:
1. Develop the idea of DARTS and implement a soft-pruning mechanism for selecting architecture operations.
2. Weights and architectures are updated simultaneously.

[Kirill Akhmetzyanov @Kirill_AI]

🎓Meta-Learning of Neural Architectures for Few-Shot Learning
by Thomas Elsken, Benedikt Staffler, Jan Hendrik Metzen, Frank Hutter

#shortreview #ML #AI #datascience #deeplearning #programming #machinelearning

[Kirill Akhmetzyanov @Kirill_AI]

Idea: combine progressive search PNAS with first-order learner Reptile.

This article is the first in the field of meta-learning + NAS. Generating dynamic architecture gives higher accuracy on the benchmark compared to static meta-learning architecture.

[Kirill Akhmetzyanov @Kirill_AI]

I continue to review articles related to my scientific work!
Today's review is about meta-NAS.

🎓Auto-Meta: Automated Gradient Based Meta Learner Search
by Jaehong Kim, Sangyeul Lee, Sungwan Kim, Moonsu Cha et al.

#ML #AI #datascience #deeplearning #programming #machinelearning

[Kirill Akhmetzyanov @Kirill_AI]

New thread of reviews! Now about meta-NAS!

These methods combine #MetaLearning and #NAS. Purpose: to get the optimal architecture and rapid transfer of weights for the current task.

#ML #AI #datascience #deeplearning #100DaysOfCode