Ansh Soni Profile
Ansh Soni

@Ansh_soni1234

Followers
112
Following
289
Media
4
Statuses
32

PhD student @ UPenn

Joined October 2021
Don't wanna be here? Send us removal request.
@Ansh_soni1234
Ansh Soni
1 month
RT @KordingLab: This administration loves other countries!.
0
4
0
@Ansh_soni1234
Ansh Soni
2 months
RT @KordingLab: Why (partial) human connectomes without molecules probably are not all that useful for AI.
Tweet media one
0
6
0
@Ansh_soni1234
Ansh Soni
3 months
RT @neocentrist: You're all wrong. Autism is caused by inflation (R^2 of 0.976). And to be clear, NGDP targeting would fix this. https://….
0
2K
0
@Ansh_soni1234
Ansh Soni
4 months
RT @ArtDeza: @saranormous Because there really isn't an AGI recipe. This is something that the rigor of academia thankfully puts perspectiv….
0
1
0
@Ansh_soni1234
Ansh Soni
4 months
RT @AToliasLab: After 7 years, thrilled to finally share our #MICrONS functional connectomics results!. We recorded activity from ~75K neur….
0
95
0
@Ansh_soni1234
Ansh Soni
9 months
RT @nacloos: ⁉️What do model-neural similarity scores tell us?. To systematically explore this for different metrics, we develop new numeri….
0
61
0
@Ansh_soni1234
Ansh Soni
9 months
RT @Ansh_soni1234: @aran_nayebi @jeffrey_bowers @RylanSchaeffer Hi @aran_nayebi, just wanted to clarify that the paper @jeffrey_bowers has….
0
3
0
@Ansh_soni1234
Ansh Soni
9 months
There is some discussion to be had about the relationships between these metrics but if we continue with this type of comparison without further introspection it becomes easy to "metric shop" until they find something that fits a specific conclusion.
1
0
6
@Ansh_soni1234
Ansh Soni
9 months
I think this is the clearest takeaway. We can see that the ranking across different models is not consistent between different measures. Here's a plot from (bit of a shameless plug) highlighting the inconsistency.
Tweet media one
@ItsNeuronal
Alex Williams
9 months
Apropros of nothing, I would like to say that we should NOT be using metrics to rank order neural network models from best to worst. We should be using metrics to make predictions, find clusters of models, etc. Triangle inequality is the antidote to Goodharting. Details 👇.
1
3
24
@Ansh_soni1234
Ansh Soni
10 months
RT @FrameworkPuter: Cramming 96GB into my laptop just to feel something.
Tweet media one
0
149
0
@Ansh_soni1234
Ansh Soni
10 months
RT @NobelPrize: The 2024 #NobelPrize laureates in chemistry Demis Hassabis and John Jumper have successfully utilised artificial intelligen….
0
4K
0
@Ansh_soni1234
Ansh Soni
10 months
RT @NobelPrize: BREAKING NEWS.The Royal Swedish Academy of Sciences has decided to award the 2024 #NobelPrize in Physics to John J. Hopfiel….
0
13K
0
@Ansh_soni1234
Ansh Soni
10 months
RT @KordingLab: How @LyleUngar and myself use LLMs for fun and profit:
0
10
0
@Ansh_soni1234
Ansh Soni
11 months
Upshot: we need to care about the metrics. First work with my new supervisor @kordinglab, my old supervisor @meenakshik93 and the inimitable @sudh8887. Was great fun to look at the basis of the standard approaches we all use. (7/7).
1
0
8
@Ansh_soni1234
Ansh Soni
11 months
Work by @nacloos Christopher Cueva and @GuangyuRobert on alignment scores relation to behaviors and work by Yeena Han and @thisismyhat about system identification within networks only works with some measures. (6/7).
1
0
8
@Ansh_soni1234
Ansh Soni
11 months
Also check work by @ermgrant and @SaxeLab on how the choice of comparison technique should be guided by the specific use case, @bkhmsi @ABosselut and @martin_schrimpf with similar results for language (5/7).
1
0
10
@Ansh_soni1234
Ansh Soni
11 months
With all the hyperparameters such as metric choice, which layer of the model is readout from, the recording modality, the dataset and much more, we caution researchers from making strong conclusions about the brain utilizing this method without further scrutiny. (4/7).
1
1
13
@Ansh_soni1234
Ansh Soni
11 months
There are many established brain network comparison metrics (e.g. RSA, Linear Predictivity, CKA, Soft Matching). Maybe unsurprisingly, I find that the conclusions drawn heavily depend on the choice of this metric (3/7)
Tweet media one
1
0
11