Maso @Masomatics

On Wed 8 May at ICLR2024, we will be presenting the recent work of Neural Fourier Transform, a general approach of equivariance/symmetry learning that naturally extends the classic Fourier transform. Please drop by to check it out! arxiv.org/abs/2305.18484

0 7 18 2K 10

View Details

Maso @Masomatics

3 years ago

@Kyoto_1192 ここは日本で一番うまいラーメンです。自分も再開店情報を心待ちにしています。

1 0 1 222 0

View Details

Yann LeCun @ylecun

3 years ago

Modeling the world for action by generating pixel is as wasteful and doomed to failure as the largely-abandoned idea of "analysis by synthesis". Decades ago, there was a big debate in ML about the relative advantages of generative methods vs discriminative methods for classification. Learning theorists, such as Vapnik, argued against generative methods, pointing out that training a generative modeling was a way more difficult than classification (from the sample complexity standpoint). Regardless, a whole community in computer vision was arguing that recognition should work by generating pixels from explanatory latent variables. At inference time, one would infer the configuration of latent variables that generated the observed pixels. The inference method would use optimization: e.g. use a 3D model of an object and try to find the pose parameters that reproduce the image. This never quite worked, and it was very slow. Later, some people converted to the Bayesian religion and tried to use Bayesian inference for the latent (e.g. using variational approximations and/or sampling). At some point, when Non-Parametric Bayes and Latent Dirichlet Allocation became the rage in text modeling, some folks heroically attempted to apply that to object recognition from images. >>> THIS WAS A COMPLETE AND UTTER FAILURE <<< If your goal is to train a world model for recognition or planning, using pixel-level prediction is a terrible idea. Generation happens to work for text because text is discrete with a finite number of symbols. Dealing with uncertainty in the prediction is easy in such settings. Dealing with prediction uncertainty in high-dimension continuous sensory inputs is simply intractable. That's why generative models for sensory inputs are doomed to failure.

112 216 2K 582K 951

View Details

Maso @Masomatics

3 years ago

@_onionesque Excited to know that there are people looking at similar things! arxiv.org/abs/2210.05972

0 0 5 160 0

View Details

Maso @Masomatics

3 years ago

ギックリ腰になったチョコラータのグリーンディに攻撃を受けているような状態(下に下がれない

0 0 0 230 0

View Details

Maso @Masomatics

3 years ago

It is my wish that it will illuminate a path toward a new direction of fun for the Symmetry learning in ML.

0 0 0 129 0

View Details

Maso @Masomatics

3 years ago

By extending our scopes from "action on a single data instance" to "action on a function on the entire data space", we can extend the theory of FT and equivariance learning to the cases involving nonlinear actions, such as nonlinearly deformed shift or Object rendering, etc

1 0 0 168 0

View Details

Maso @Masomatics

3 years ago

Ever got frustrated by the fact that most application/theory of Symmetry in ML are built on "linear action" on data? We just recently released a new preprint to rethink the equivariance learning by establishing the "Neural" extension of Fourier Transform. arxiv.org/abs/2305.18484

1 3 3 1K 2

View Details

Stat.ML Papers @StatMLPapers

3 years ago

Neural Fourier Transform: A General Approach to Equivariant Representation Learning. (arXiv:2305.18484v1 [stat.ML]) ift.tt/BJXMoLw

0 40 193 32K 88

View Details

Maso @Masomatics

4 years ago

For some reason my wife, not being so familiar with English names, thought that the person I strive to be is Mike "Ty"son. Not that it makes so much difference to me in terms of sheer Badassness, but I wonder how "Orri"son became "Ty"son in her mind LOL

0 0 0 110 0

View Details

Maso @Masomatics

4 years ago

@sakurako_iroha これ死ぬほどわかります　「ほっとくと妻にやらせてしまう」という動機「のみ」が自分の家事行動を動かしてる

1 0 1 80 0

View Details

Maso @Masomatics

4 years ago

@kyfwfw そうか、、、Consciousnessとかされてる方なんですね！！SNN使ったことないですが博士時代はPoisson Processずっとやっていたので益々ご縁感じます！　 MSNに飛ぶ機会は多分暫くないですね、、、まだ東京中心になりそうです脳といえば、ポスドク時代は京大の石井信先生には大変お世話になりました。