Abstract
We study the collaborative PAC learning problem recently proposed in Blum et al. [3], in which we have k players and they want to learn a target function collaboratively, such that the learned function approximates the target function well on all players' distributions simultaneously. The quality of the collaborative learning algorithm is measured by the ratio between the sample complexity of the algorithm and that of the learning algorithm for a single distribution (called the overhead). We obtain a collaborative learning algorithm with overhead O(ln k), improving the one with overhead O(ln2 k) in [3]. We also show that an Ω(ln k) overhead is inevitable when k is polynomial bounded by the VC dimension of the hypothesis class. Finally, our experimental study has demonstrated the superiority of our algorithm compared with the one in Blum et al. [3] on real-world datasets.
| Original language | English (US) |
|---|---|
| Pages (from-to) | 3598-3607 |
| Number of pages | 10 |
| Journal | Advances in Neural Information Processing Systems |
| Volume | 2018-December |
| State | Published - Jan 1 2018 |
| Event | 32nd Conference on Neural Information Processing Systems, NeurIPS 2018 - Montreal, Canada Duration: Dec 2 2018 → Dec 8 2018 |
Fingerprint
ASJC Scopus subject areas
- Computer Networks and Communications
- Information Systems
- Signal Processing
Cite this
Tight bounds for collaborative PAC learning via multiplicative weights. / Chen, Jiecao; Zhang, Qin; Zhou, Yuan.
In: Advances in Neural Information Processing Systems, Vol. 2018-December, 01.01.2018, p. 3598-3607.Research output: Contribution to journal › Conference article
}
TY - JOUR
T1 - Tight bounds for collaborative PAC learning via multiplicative weights
AU - Chen, Jiecao
AU - Zhang, Qin
AU - Zhou, Yuan
PY - 2018/1/1
Y1 - 2018/1/1
N2 - We study the collaborative PAC learning problem recently proposed in Blum et al. [3], in which we have k players and they want to learn a target function collaboratively, such that the learned function approximates the target function well on all players' distributions simultaneously. The quality of the collaborative learning algorithm is measured by the ratio between the sample complexity of the algorithm and that of the learning algorithm for a single distribution (called the overhead). We obtain a collaborative learning algorithm with overhead O(ln k), improving the one with overhead O(ln2 k) in [3]. We also show that an Ω(ln k) overhead is inevitable when k is polynomial bounded by the VC dimension of the hypothesis class. Finally, our experimental study has demonstrated the superiority of our algorithm compared with the one in Blum et al. [3] on real-world datasets.
AB - We study the collaborative PAC learning problem recently proposed in Blum et al. [3], in which we have k players and they want to learn a target function collaboratively, such that the learned function approximates the target function well on all players' distributions simultaneously. The quality of the collaborative learning algorithm is measured by the ratio between the sample complexity of the algorithm and that of the learning algorithm for a single distribution (called the overhead). We obtain a collaborative learning algorithm with overhead O(ln k), improving the one with overhead O(ln2 k) in [3]. We also show that an Ω(ln k) overhead is inevitable when k is polynomial bounded by the VC dimension of the hypothesis class. Finally, our experimental study has demonstrated the superiority of our algorithm compared with the one in Blum et al. [3] on real-world datasets.
UR - http://www.scopus.com/inward/record.url?scp=85064839438&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85064839438&partnerID=8YFLogxK
M3 - Conference article
AN - SCOPUS:85064839438
VL - 2018-December
SP - 3598
EP - 3607
JO - Advances in Neural Information Processing Systems
JF - Advances in Neural Information Processing Systems
SN - 1049-5258
ER -