CN-121998037-A - Semantic feedback guided structure self-adaptive federal learning method, system and medium

CN121998037ACN 121998037 ACN121998037 ACN 121998037ACN-121998037-A

Abstract

The invention discloses a semantic feedback guided structure self-adaptive federal learning method, a system and a medium, which belong to the technical field of federal learning and deep neural network training, wherein the system comprises FedBR a basic framework, a multi-source feedback statistical module, a block scheduling strategy module and a coverage fairness constraint module. The invention introduces a probability type deep block scheduling strategy, so that the system can gradually deviate to a block combination with better effect according to accumulated feedback, and a certain degree of online self-adaption capability is shown in an experimental scene. The non-stationary feedback smoothing mechanism of time window plus exponential sliding average is adopted, so that the influence of single-round noise and partial observation on strategy updating is reduced, the strategy updating is more stable, and the occurrence frequency of strategy concussion and extreme paranoid conditions is reduced. When the probability distribution P is updated, probability distribution smoothing and coverage fairness constraint is introduced, so that extreme cases of deep block which is not selected for a long time are obviously reduced, and the overall training participation of the deep block is improved.

Inventors

AI BING
HU ZEKUN
ZHANG MEILING
WANG XIAOFENG
CHEN SHENGBING
Min Chengzhi

Assignees

合肥大学

Dates

Publication Date: 20260508
Application Date: 20260120

Claims (10)

1. A semantic feedback guided structure self-adaptive federal learning method is characterized by comprising the following specific steps: Step one, model partitioning; Step two, evaluating the client capability and determining a block number candidate interval; step three, decision of block number budget alpha (u, t); Step four, full model training round judgment and issuing block set generation, wherein full (t) is made to represent whether a t-th round is a full model training round or not, and when full (t) =1, the server issues all B blocks to the client; fifth, probability type deep block selection and issuing When full (t) =0, the server selects K_deep (u, t) blocks from the deep block set S_deep to be issued to the client u, maintains four types of state quantities for each deep block_b, namely, a dominant mean value or smooth gain estimation mu (b) of the block, an uncertainty variance sigma2 (b) of the block, a coverage fairness constraint weight lambda (b), the number of rounds gap (b) undergone by the block since last selection, sets a variance lower limit sigmafloor, a time window W, an old compensation coefficient stale _gamma and a compensation upper limit stale _cap, and firstly samples disturbance xi (b) from a Gaussian distribution with a mean value of 0 and a variance of not less than sigmafloor2 for each round to construct a random dominant sample: theta(b)=mu(b)+xi(b)+stale_gamma min(gap(b)/W,stale_cap) and overlapping the coverage constraint items to obtain a sequencing priority score: s(b)=theta(b)+lambda(b) The server sorts the deep blocks in descending order according to S (b), and selects the front K_deep (u, t) deep blocks with highest scores under the budget K_deep (u, t) to form a set A (u, t) to be sent to the client; Step six, client block wise local training; step seven, returning and aggregating with a server; step eight, multi-source feedback statistics and time window smoothing, wherein the server counts real rewards real_forward and proxy rewards proxy_forward in a time window with the length of W for each deep block, and discounts historical feedback in the window by using exponential sliding average to obtain smooth feedback estimation r_hat (b, t), wherein a smoothing coefficient beta is used for distributing weight between current round feedback and historical feedback; Step nine, updating probability distribution P by a cover fairness constraint module through probability smoothing and cover fairness constraint; The server updates state quantity mu (b) and sigma2 (b) of a deep block according to r_hat (b, t) and updates coverage constraint weight lambda (b), maintains a selection history of the latest W rounds for each deep block, calculates recent coverage f (b) of the block, sets a target coverage level c_target by a system, and gradually increases lambda (b) if the coverage of a block is lower than c_target for a long time: lambda(b)=clip(lambda(b)+cov_lr(c_target-f(b)),0,lambda_max) Wherein cov _lr is a smaller learning rate, lambda_max is an upper bound, clip represents a clipping function, and meanwhile, if not, the gap (b) is updated according to the selection, if not, the gap (b) is added with one, and if so, the gap (b) is cleared; Step ten, bandit, feeding back construction and block number budget state updating; Step eleven, entering the next round.
2. The semantic feedback guided structured adaptive federal learning method according to claim 1, wherein in step one, the backbone network is divided into B structure blocks in forward order, each block being composed of several consecutive layers.
3. The semantic feedback guided structure adaptive federation learning method according to claim 2, wherein in the second step, the server selects a client set U (t) participating in training from all clients in the t-th round, for each selected client U, the server calculates three normalized capability scores, namely a calculation capability h_comp (U), an available memory h_mem (U) and an uplink bandwidth h_comm (U), the value ranges from 0 to 1, and thus obtains a comprehensive capability score C (U): C(u)=0.5h_comp(u)+0.3h_mem(u)+0.2h_comm(u) The comprehensive ability score C (u) is cut and normalized to fall between 0 and 1; The server sets two thresholds tau1 and tau2 so as to divide the client into three gears of a weak end, a middle end and a strong end, and configures an optional block number interval [ k_min (u), k_max (u) ] of the client u; Let k_min (u) =1, k_max (u) =min (2, k_max) when C (u) is less than or equal to tau1, let k_min (u) =min (2, k_max) when tau1 is less than C (u) and C (u) is less than tau2, let k_min (u) =min (3, k_max), when C (u) is greater than or equal to tau2, let k_min (u) =min (3, k_max), k_max (u) =min (4, k_max), where k_max is the global maximum number of blocks allowed for a single round.
4. The method for learning the adaptive federation of semantic feedback guided architecture according to claim 3, wherein in the third step, at the beginning of each round of training, the server randomly selects an integer for the client u within the interval [ k_min (u), k_max (u) ] as the block number budget alpha (u, t) of the round, the server maintains two state quantities for each client u, p_mean (u, m) represents the historical feedback mean value of the number of selected blocks m, n_count (u, m) represents the cumulative number of the number of selected blocks m, and the t-th round calculates V (u, m, t) for each candidate m: V(u,m,t)=p_mean(u,m)+sqrt(ln(t+1)/(n_count(u,m)+1)) V (u, m, t) is the decision value of the t-th round for comparing the different candidate block numbers m, let alpha (u, t) take m which maximizes V (u, m, t).
5. The semantic feedback guided architecture adaptive federal learning method according to claim 4, wherein in step four, the number of deep selected blocks is defined as: K_deep(u,t)=alpha(u,t)-|S_fix| where S_fix represents the number of shallow necessary blocks, and if K_deep (u, t) is less than 0, let it be 0.
6. The method for adaptive federal learning based on semantic feedback guidance according to claim 5, wherein in step six, the client u receives a block parameter set issued by the server, receives full model parameters when full (t) =1, receives s_fix union a (u, t) when full (t) =0, and trains several training rounds on the local data.
7. The method according to claim 6, wherein in step nine, when updating the probability distribution, the server normalizes the score s (b) by temperature softmax to obtain a base distribution p_base (b, t), the temperature parameter tau_soft takes 1.0 to 2.0 to avoid excessively sharp distribution, then smoothly merges and normalizes the p_base (b, t) and the previous round of distribution P (b, t-1) by step eta_p to obtain a probability distribution P (b, t), then mixes the probability distribution P (b, t) with the discovery rate eps (t), specifically constructs an even distribution U on the deep block set, and makes p_new (b) = (1-eps (t))p (b, t) +eps (t) ×u (b) for each deep block set to obtain a new probability distribution p_new.
8. The method of claim 7, wherein the step of ten, bandit feedback building and block number budget state updating comprises the steps of the server constructing a scalar feedback r_band (u, t) for updating the block number budget bandit for each client u, recording the selected candidates of the round as m_star, i.e., m_star=alpha (u, t), then updating only the state quantity corresponding to m_star by the server, updating p_mean (u, m_star) with an exponential moving average, and adding n_count (u, m_star) by one; Step eleven, the next round is entered, namely the server uses the updated global model parameters, the deep probability distribution P_new and p_mean (u, m) and n_count (u, m), and the t+1st round of training is entered.
9. An adaptive federal learning system for use in the method of any of claims 1-8, the system comprising FedBR a base frame, a multi-source feedback statistics module, a block scheduling policy module, and an overlay fairness constraint module.
10. A computer storage medium for storing program data which, when executed by a computer, is adapted to implement the semantic feedback guided architecture adaptive federal learning method of any one of claims 1-8.

Description

Semantic feedback guided structure self-adaptive federal learning method, system and medium Technical Field The invention relates to the technical field of federal learning and deep neural network training, in particular to a semantic feedback guided structure self-adaptive federal learning method, system and medium. Background The traditional federal learning method, such as FedAvg, issues a complete global model to the participating clients during each round of training, and aggregates and updates the complete parameters after the local training of the clients. The method has good performance in the scene of relatively independent and same distribution of data and small difference of equipment capacity, but in the actual system with large difference of terminal calculation power and bandwidth, the complete model synchronization brings large communication burden, and weak-end equipment is difficult to stably participate for a long time. To alleviate the above problems, a class LAYER WISE or block wise federal learning approach has emerged in recent years that reduces communication overhead by training and returning only a portion of the hierarchy or portion of the blocks of the model, and enables weaker clients to participate in training on a smaller subset of the model. Existing work has typically employed static or heuristic deterministic strategies to select these partial parameters. FedBR proposes a representative block wise federal training framework. The main thought comprises the following steps: a. the depth network is divided into a plurality of consecutive blocks in forward order. B. And the server determines the block number k of the training round of the client according to the computing power and the communication capability of the client. C. After k is determined, continuously issuing k deep blocks at the tail end to the client, and forming a continuous tail section by adopting a fixed issuing mode. D. The client only receives alpha global blocks with continuous tail ends, and combines the alpha global blocks with the rest blocks reserved locally to form a complete local model, the complete local model is returned after the training is completed locally, and the server executes aggregation update on the complete model. E. The client training goal is introduced with block regularization (block wise regularization) and block knowledge distillation (block wise knowledge distillation) for reducing locality Xi Pianzhi brought by dependent co-distributed data. F. Full model training rounds (full round) are performed once every few rounds, and full blocks are issued for full model training to recalibrate global alignment and rewards scale. FedBR achieves better effects in communication compression and heterogeneous adaptation, and is an important baseline for block wise federal learning. However, in the scene of strong independence and large difference between the same distribution and the terminal capability, a plurality of problems still exist: 1. FedBR, after FedBR gives the budget k of the number of blocks of the client, the server determines the deep block sub-segment by means of 'continuous k blocks at the tail end are fixedly issued', and dynamic adjustment is difficult to achieve according to the change of the contribution of each block in the training stage, and the issuing rule is relatively rigid. 2. Under the condition that only part of blocks participate in feedback, the server has limited observation on the block effect, and the server is easy to generate paranoid and concussion directly based on a single-round income updating strategy. 3. Due to feedback paranoid and strong end dominance, part of deep blocks may be selected for a long time with low probability, and even be simply regarded as 'poor-effect blocks' to be continuously weakened, so that the phenomenon of block starvation occurs. Therefore, on the premise of keeping FedBR main body pipelines unchanged, it is necessary to perform policy level upgrading aiming at the block scheduling and feedback updating sub-module, so that the scheduling has certain self-adaptive capacity, the feedback is smoother and more stable, and meanwhile, a re-exploration opportunity is provided for the underestimated block in the long-term training process. Based on the above, the invention designs a semantic feedback guided structure self-adaptive federal learning method, a system and a medium to solve the above problems. Disclosure of Invention Aiming at the defects existing in the prior art, the invention provides a semantic feedback guided structure self-adaptive federal learning method, a system and a medium. On the premise of keeping the basic block wise pipeline of FedBR unchanged, the invention carries out strategy layer upgrading on a block scheduling and feedback updating sub-module, and forms a closed-loop mechanism which takes exploration utilization, stability and coverage into consideration by constructing block-level semantic feedback