Channel
@one_figure
On this record: Growth · Engagement · Posts · Cite this entry
43subscribers
-1 since we began measuring on 7 August 2026
Risers and fallers across the register · movement among entries of Under 1,000.
| Telegram ID | -1001585135644 |
|---|---|
| Type | Channel |
| Username | @one_figure |
| Created | 5 July 2021 — measured — cross-checked against a third-party dataset (ext.tg_channel) |
| First recorded | 10 August 2026 |
| Last confirmed live | 24 August 2026 |
| Measurements held | 3 |
| Confirmed unchanged | 1 time, most recently 24 August 2026 |
| On Telegram | t.me/one_figure |
| Measured (UTC) | Subscribers | Change |
|---|---|---|
| 24 Aug 2026, 04:35 | 43 | -1 |
| 10 Aug 2026, 16:00 | 44 | no change |
| 7 Aug 2026, 19:56 | 44 | first reading |
15 posts held, back to 5 July 2021 — the reader has not yet reached the start of this channel’s public history, so older posts may sit further back, unread. Read across 1 pageof Telegram’s post history, 20 posts per page.
Nothing published in the last 30 days. ERR and ER are rolling 30-day measures, so there is nothing to compute — we hold 15 posts for this entry, the most recent from 3 November 2021. An engagement rate over an empty window would be a number about nothing.
When training DNNs on datasets with noisy labels, the model quickly learns the clean labels. After such successful warm-up training, the model produces large losses on noisy samples and small losses on clean labels. This allows LNL methods to identify noisy samples and discard them, or deal with them in another way. However, if trained longer the model begins to memorize noise and no LNL method can provide any benefi…
If your dataset has 5, 10 or even 100 noisy labels per each clean label, deep neural networks will not ocerfit to the noise. Given that you have enough clean labels. Large batch sizes, large networks and small learning rates increase robustness to label noise. Deep learning is robust to massive label noise, Rolnick et. al, 2018, Arxiv. https://arxiv.org/abs/1705.10694
An agent shows tool use in a zero-shot setting. "Finding better agents and improving the quality of evaluation are the same problem" Deepmind achieves crazy Reinforcement Learning results by: 1. Dynamically generating games. It is a smooth space of a myriad of possible games, from trivial to highly complex, from cooperative to competitive, from multi-agent to single-player, from balanced to completely dishonest. 2.…
KagNet parses concepts from a question-answer pair, builds a knowledge graph of relations between concepts, and uses attention to weight the strength of relations in this question context. Finally, it picks the answer with the strongest relation. KagNet attempts commonsense reasoning by combining information from a knowledge graph and BERT embeddings of a question-answer pair. By using the ConceptNet knowledge grap…
Extracting representations not from the layer you use for training, but from the previous one
Authors of SimCLR, the self-supervised SOTA for the moment, compare representations obtained from different layers of the output head. Surprise: taking the final output is not the best approach! In SimCLR they feed the output of the AvgPool layer of a backbone into an MLP: z = g(h). The network is trained on the outputs g(h) of this MLP, but they evaluate the network on the inputs h of the MLP. See picture below. T…
A clear view
Comparing semi-supervised learning approaches on a toy problem. Black and white dots are few labelled examples, grey dots are unlabelled and numerous. The decision boundary when using only labelled examples (dashed line) is drastically different from the true decision boundary. SSL uses unlabelled data to get better models. Authors compare SSL approaches in a fair setting. They absolutely destroy SSL papers: * Turns…
The effectiveness of government measures against COVID-19 spread computed from 41 countries, mostly European. Wow, the paper has like 50 pages of appendix material. The effectiveness of eight nonpharmaceutical interventions against COVID-19 in 41 countries, Brauner et al., 2020 Found via this ACT post.
Models pre-trained on ImageNet perform poorly when recognizing objects in unusual contexts. This issue goes away if we train large models on larger datasets, as authors of Big Transfer (BiT) paper demonstrate. BiT is a procedure to pre-train and fine-tune models for better performance on downstream tasks. Key pre-training tricks: * GroupNorm + WeightStandartization instead of BatchNorm * Simple augmentations: random…
Channel name was changed to «Figure 1»
Channel photo updated
Showing the 12 most recent of 15 posts we hold for @one_figure. View and reaction counts are the latest single reading for each post, not a live figure, and a recent post is still accumulating both. A view count marked ≈ was rounded by Telegram before we ever saw it — t.me prints views in full below 1,000 and to three significant figures above, so ≈1,200,000 means somewhere between 1,150,000 and 1,249,999. Unmarked counts are exact. Text is reproduced from the public post preview and truncated for length.
A live page changes as we take new readings, so a citation should name the measurement it is based on, not just the URL. The line below cites the subscriber count as measured 24 August 2026 — this entry's latest reading, not the date you are reading this.
“Figure 1” (@one_figure), 43 subscribers as measured 24 August 2026. Telegram Register, tgregister.com/channel/one_figure.
Full measurement history, CC BY 4.0. Every reading this register holds for this entry, not just the latest one, as a dated, downloadable record: CSV · JSON. Free to use with attribution to tgregister.com. Each file carries its own generation timestamp, which is the figure to cite for exactly when the data was retrieved.