ABCDEFGHIJKLMNOPQRSTUVWXYZ
1
SnoBlog TitleBlog LinkKesav's SummarySurya's SummaryLikes/DislikesBlog Source
2
1Enabling Continual Learning in Neural Networkshttps://deepmind.com/blog/enabling-continual-learning-in-neural-networks/Consolidation helps in retaining information, let it be impriting memories from quick-learning parts of the brain to slow-learning parts by recalling (or) avoiding completely overwriting previously learnt connections while learning new tasks. The aforementioned two types are termed as systems consolidation and synaptic consolidation respectively.

This blog discusses about catastrophic forgetting in neural networks where they tend to forget the connections learnt on previous tasks while learning new tasks. In synaptic consolidation, our brain tends to overwrite the connections based on how important they were in previous tasks. In the same way, catastrophic forgetting is addressed by protecting those connections which are still relevant in the old tasks by assigning some importance to each connection. If each connection is considered a spring, the protection is proportional to its stiffness. A new algorithm called Elastic Weight Consolidation is designed to overcome this problem in neural networks.

Overcoming this problem helps in multi-tasking and hence gives rise to the possibility of AGI.		Tries to solve catastrophic forgetting using neuro-inspired theories. After learning a task, the most important connections are found (how? not sure) and locked (I'm using words loosely). Was tested by trying to learn multiple Atari games sequentially (difficult problem because each game requires it's own strategy, or we can look at this as a single task with non-iid samples and online learning). Typical networks fail because of the non-iid nature, while the propsed network locks weights to remember learning from older samples.


My views: DeepMind's blog summaries catastrophic forgetting as a problem when you're trying to learn new tasks. I feel this definition is narrow. In the context of RL, it happens even when you're learning just one task too. (Example: I had this problem while training A2C on CartPole) In this case, the problem occurs because of having non-iid samples while training the network. So I think batch RL approaches wouldn't suffer with this problem, but this is based on my intuition. Need to check it with people who know more. I like Wikipedia's definition more, "the tendency of an artificial neural network to completely and abruptly forget previously learned information upon learning new information." 		Kesav:
Likes:
It is interesting to see three inter-disciplinary areas such as physics, deep learning and cognitive science in addressing one problem. The blog is simple, so, it became easy to understand it well.

Dislikes:
No reason to hate it.
DeepMind
3
2OpenAI Fivehttps://openai.com/blog/openai-five/OpenAI
4
3Graph Powered ML at Googlehttps://ai.googleblog.com/2016/10/graph-powered-machine-learning-at-google.htmlGraph powered machine learning has many interesting applications at Google such as reminders and smart reply in Inbox, in Google Photos, etc. Learning can be performed in different ways and graph based semi-supervised learning is inspired from how humans learn i.e. learn to generalize and then bridge the gap between what they already know. Graph based semi-supervised learning is advantageous because 1. It jointly learns the underlying distribution of data from labeled and unlabeled data at the same time unlike any other learning techniques 2. It combines different types of information into a single graph representation for learning.

Typical learning workflow in this algorithm:
1. Create a graph from the data that we have where edges represent relationships between the nodes/objects. Few of the nodes are labeles while the rest are unlabeled.
2. Learn a label for each nodes from its neighbors while taking into account of the strength of the connection i.e. edge weight.
3. Repeat step-2 for few iterations		Typically, semi-supervised approaches train on labeled data and use that model on unlabeled data. In this graph-based semi-supervised approach, the joint of labeled and unlabeled data is modelled at once.

Step 1: Construct a graph from data with nodes representing datapoints (loosely) and edges representing similarities. (how? that has not been mentioned in the blog)

Step 2: Label all nodes by propagating information to unlabeled nodes. (done using an iterative approach employing ideas from convex optimization, called streaming algorithm)

I believe (this has not been mentioned in the article) these new labels for unlabeled data combined with the labeled dataset is used for training a ML model.		Kesav:
Likes:
Applications are interesting. I came to know about an interesting team at Google which might be helpful for me in exploring teams after joining.

Dislikes:
I knew about semi-supervised graph learning already and there is no technical enlightenment received except for the applications part.
Google AI
5
4
6
5
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100