E0234 PPT
E0234 PPT
E0234 PPT
DECISION
Harsh Vishwakarma
21532
MTech, CSA
Introduction
Applications of JL lemma
Speeding up JL lemma
Table of Contents
Introduction
Applications of JL lemma
Speeding up JL lemma
Introduction
What is interpretability
I A model’s ability to provide insights for its decisions or inner
working, whether intrinsically or not, is referred to as
interpretability.
I Complex models, such as transformers, cannot provide
interpretations out of the box, and therefore posthoc
techniques are typically applied. The representations of an
interpretation include, among others, rules, heatmaps, and
feature importance.
Interpretability of Transformer
Introduction
Applications of JL lemma
Speeding up JL lemma
Optimus Transformer Interpretibility
Proof:
kv k22 ≤ (1 − )kuk22 k
P
kuk22 x
2
= ≤ (1 − )kuk2
p
=[x ≤ (1 − )p]
h i
= e −λx ≥ e −λ(1−)p ( for all λ ≥ 0)
By Markov’s inequality
E [x]
P[x ≥ a] ≤
a
Lemma 3
E e −λx
h i
−λx −λ(1−)p
P e ≥e ≤ −λ(1−)p
e h i
2
p
Y E e −λxi
≤ (as xi ’s are i.i.d.)
e −λ(1−)p
i=1
h 2
i p
E e −λxi
≤ −λ(1−)
e
p
1
≤ √
1 + 2λ · e −λ(1−)
( using the m.g.f. of the χ2 distribution )
Lemma 3
put λ = (optimal value after differentiating )
2(1 − )
≤ [(1 − )e ]p/2
Using inequality log(1 − x) < −x − x 2 /2 and putting value of p,
≤ n−2
Combining the above lemmas we get,
2
P kv k22 ∈/ (1 − )kuk22 , (1 + )kuk22 ≤ 2
n
Now the above result is for any two pairs of points. We can use
Union Bound for all such pairs of points.
(n2)
X
P (Bounds fail for any pair of points ) ≤ P (Bounds fail for xi , xj )
i=1
n(n − 1) 2 1
≤ 2
≤1−
2 n n
Table of Contents
Introduction
Applications of JL lemma
Speeding up JL lemma
Applications
2 Step process:
Step 2 : SVD of Y
p
X
Y = AΛB T = λi ai biT
i=1
` ` kf (x i ) − f (x j )kH
if ≤ kx i − x j kH ≤ =⇒ (1−)pξ ≤ < (1+)pξ
4 2 kx i − x j kH
` `
if kx i − x j kH > =⇒ kf (x i ) − f (x j )kH > (1 − )pξ
2 2
forall
4
xi, xj ∈ Zd2 , with probability at least 1 − e −c p where c > 0.
Approximate Nearest Neighbor Search using random
projection
Introduction
Applications of JL lemma
Speeding up JL lemma
Speeding up the JL transform