Continuous Learning with Deep Architectures

Vincenzo Lomonaco
University of Bologna

• Deep Learning state-of-the-art performances
• Mainly supervised training with huge and fixed datasets.

# of possible 227x227 RGB images
3,9 ∙ 10372282

(Hence ubiquitousness and autonomy)

𝑡
𝑋1 𝑌1
𝑓𝜃: 𝑋1 → 𝑌1

𝑡
𝑋2 𝑌2𝑓𝜃: 𝑋1 ∪ 𝑋2 → 𝑌1 ∪ 𝑌2

• Higher and realistic time-scale where data (and tasks)
become available only during time
• No access to previously encountered data.
• Constant computational and memory resources.
• Incremental development of ever more complex
knowledge and skills.

Dataset, Benchmark, code and additional
information freely available at:
vlomonaco.github.io/core50
LomonacoV. and Maltoni D. CORe50: a New Dataset and Benchmark for ContinuousObject Recognition.CoRL2017.

Architectural
Regularization Reharshal
CWR PNN
EWC
SI
LWF
ICARL
GEM

𝑊𝑡𝑊𝑏...
𝐿0, … , 𝐿 𝑛−1 𝐿 𝑛
?
?
?
?

𝑊𝑡𝑊𝑏...
𝐿0, … , 𝐿 𝑛−1 𝐿 𝑛
?
?

𝑊𝑡𝑊𝑏...
𝐿0, … , 𝐿 𝑛−1 𝐿 𝑛
?
?
𝑊𝑐
𝐿 𝑛

𝑊𝑡𝑊𝑏...
𝐿0, … , 𝐿 𝑛−1 𝐿 𝑛
𝑊𝑐
𝐿 𝑛

𝑊𝑡𝑊𝑏
...
𝐿0, … , 𝐿 𝑛−1 𝐿 𝑛
𝐿 𝜇 = 𝐿 𝜇 + 𝜆
𝑘
Ω 𝑘
𝜇
𝜃 𝑘 − 𝜃 𝑘
2
𝑤 𝑘
𝑣
=
𝑡 𝑣−1
𝑡 𝑣
𝜕𝐿
𝜕𝜃 𝑘
∙
𝜕𝜃 𝑘
𝜕𝑡

Combining Architectural and Regularization
approaches
LomonacoV. and Maltoni D. Continuous Learning in Single-Incremental-TaskScenarios.To be published.

Architectural
Regularization Reharshal
CWR PNN
EWC
SI
LWF
ICARL
GEM
AR1

𝑊𝑡 𝑊𝑐𝑊𝑏...
𝐿0, … , 𝐿 𝑛−1 𝐿 𝑛 𝐿 𝑛

𝑊𝑡𝑊𝑏...
𝐿0, … , 𝐿 𝑛−1 𝐿 𝑛
𝐿 𝜇 = 𝐿 𝜇 + 𝜆
𝑘
Ω 𝑘
𝜇
𝜃 𝑘 − 𝜃 𝑘
2
𝑤 𝑘
𝑣
=
𝑡 𝜇−1
𝑡 𝜇
𝜕𝐿
𝜕𝜃 𝑘
∙
𝜕𝜃 𝑘
𝜕𝑡

𝑊𝑡𝑊𝑏...
𝐿0, … , 𝐿 𝑛−1 𝐿 𝑛
• Computational efficient
(independent from the
number of training
batches)
• Just one Ω 𝑘(running sum
+ max clip)
• CWR with zero-init
• CWR with mean-shift

0
5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
1 2 3 4 5 6 7 8 9
Naive
LwF
EwC
Syn
CwR
AR-1
Cumulative

1. Continuous Unsupervised Learning
2. Continuous Reinforcement Learning

Continuous Learning with Deep Architectures

More Related Content

Similar to Continuous Learning with Deep Architectures

More from Vincenzo Lomonaco

Recently uploaded

Continuous Learning with Deep Architectures

Editor's Notes